Content uploaded by Mohammadnoor Ahmad Mohammad Injadat
Author content
All content in this area was uploaded by Mohammadnoor Ahmad Mohammad Injadat on May 21, 2021
Content may be subject to copyright.
Artificial Intelligence Review manuscript No.
(will be inserted by the editor)
Machine Learning Towards Intelligent Systems: Applications,
Challenges, and Opportunities
MohammadNoor Injadat ·Abdallah Moubayed ·
Ali Bou Nassif ·Abdallah Shami
Received: date /Accepted: date
Abstract The emergence and continued reliance on the Internet and related technologies
has resulted in the generation of large amounts of data that can be made available for anal-
yses. However, humans do not possess the cognitive capabilities to understand such large
amounts of data. Machine learning (ML) provides a mechanism for humans to process large
amounts of data, gain insights about the behavior of the data, and make more informed de-
cision based on the resulting analysis. ML has applications in various fields. This review
focuses on some of the fields and applications such as education, healthcare, network secu-
rity, banking and finance, and social media. Within these fields, there are multiple unique
challenges that exist. However, ML can provide solutions to these challenges, as well as
create further research opportunities. Accordingly, this work surveys some of the challenges
facing the aforementioned fields and presents some of the previous literature works that
tackled them. Moreover, it suggests several research opportunities that benefit from the use
of ML to address these challenges.
Keywords Machine learning ·Data Analytics ·Application fields ·Research Opportunities
1 Introduction
The rapid growth of the Internet and related technologies has provided individuals, organi-
zations, and society with the opportunity to collect large amounts of data (Van Der Aalst,
MohammadNoor Injadat, Abdallah Moubayed, Abdallah Shami
Electrical & Computer Engineering Dept.
University of Western Ontario
London, ON, Canada
E-mail: minjadat@uwo.ca, amoubaye@uwo.ca, abdallah.shami@uwo.ca
Ali Bou Nassif
Computer Engineering Dept.
University of Sharjah, Sharjah, UAE
and
Electrical & Computer Engineering Dept.
University of Western Ontario
London, ON, Canada
E-mail: anassif@sharjah.ac.ae
arXiv:2101.03655v1 [cs.LG] 11 Jan 2021
2 MohammadNoor Injadat et al.
2016). However, these large amounts of data often lead to information overload. Informa-
tion overload occurs when the amount of input (e.g. data) that a human is trying to process
exceeds their cognitive capacities (Halford et al., 2005). Information overload can lead to hu-
mans ignoring, overlooking, or misinterpreting crucial information (Caban and Gotz, 2015).
Humans do not have the cognitive capacity to process large amounts of data. Therefore,
the discipline of data science has emerged. Data science combines the classic disciplines
of statistics, data mining, databases, and distributed systems in order to extract information
from large sets of data (Van Der Aalst, 2016). One approach of data analysis that data sci-
entists can implement is machine learning (ML) (Moubayed, 2018). ML allows computers
to learn without being explicitly programmed. Once the computer learns patterns from a
training set of data, it can apply what it has learned to find these patterns in similar data
(Kearns et al., 1994). Furthermore, ML allows computer systems to adapt and learn from
their experience (Wilson and Keil, 2001; Mitchell, 1997).
ML algorithms have a lot of applications. Examples are house pricing prediction, spam
filtering, education, structuring of data in healthcare systems, drug response prediction, dia-
betes research, network security, banking and finance, and social media. This work aims to
provide a brief literature review of the challenges facing different fields such as education,
healthcare, network security, banking and finance, and social media. Moreover, it presents
several research opportunities on the role and potential of using ML to address these chal-
lenges. Hence, the contributions of this paper are summarized as follows:
–Describing briefly the different challenges facing a variety of modern fields including
education, healthcare, network security, banking and finance, and social media.
–Presenting some of the previous literature works that addressed these challenges and
their shortcomings.
–Discussing the role and potential of ML in addressing these challenges and presents
potential frameworks for its deployment.
The remainder of this paper is organized as follows: Section 2 presents some of the recent
trends concerning the development and deployment of ML algorithms. Section 3 discusses
the education field. Section 4 focuses on the healthcare system and field. Section 5 presents
the challenges in network security and the potential role of ML in addressing these chal-
lenges. Section 6 sheds light on the banking and finance sector. Section 7 focuses on the
area of social media. Finally, Section 8 concludes the paper.
2 Recent Trends in Machine Learning
ML has become an extremely popular topic within development organizations that are
looking to adopt a data-driven approach to improve their business by gaining useful infor-
mation from the data they collect. With ML models, organizations can continually predict
changes in their business and make decisions accordingly. ML uses algorithms that itera-
tively learn from data to improve, describe data, and predict outcomes. Once an ML model
has been trained, it can predict new data that is given as input. The output given by the model
on the new data will depend on the data used to train the model.
The emerging growth of ML adoption in various fields is emphasized by the amount
of financial resources being allocated to deploy ML models. As illustrated in Figure 1,
the global ML market is expected to reach close to $42.5 billion CAD by the year 2024
(Columbus, 2020). Furthermore, as per McKinsey & Company’s “Notes from the AI Fron-
tier, Tackling Europe’s Gap in Digital and AI” discussion paper, the ML market could boost
Machine Learning Towards Intelligent Systems: Applications, Challenges, and Opportunities 3
Fig. 1: Global ML Forecast Columbus (2020)
economic activity growth throughout the EU by as much as 20% by the year 2030 (Bughin
et al., 2019). Moreover, the World Economic Forum predicted that a net of 58 million jobs
will be created in the coming years due to ML technologies (Forum, 2018). This highlights
the importance and positive potential impact that ML will have on the global economic mar-
ket.
Within the area of ML, one promising paradigm to adopt is federated learning (FL).
FL is a ML paradigm in which a high-quality centralized model is trained using data that is
distributed over a large number of locations. The term was first coined by Google in 2016 in
which they proposed a mechanism in which data at each location is used to independently
compute an update of the current ML model. This update is then communicated back to
a central service that aggregates these updates to compute a new global model that is dis-
tributed back to the different locations (Konecn`
y et al., 2016). Accordingly, this paradigm
adopts the “bringing the code to the data” philosophy rather than “bringing the data to the
code”. As such, the FL paradigm addresses concerns regarding the data privacy, ownership,
and locality (Bonawitz et al., 2019) (Yang et al., 2019b). Given the distributed nature of wa-
ter leak monitoring systems with sensors collecting data at various geographical locations,
FL promises to be a viable solution for extracting meaningful information from the collected
data while still maintaining its privacy and locality.
The continued projected market growth of ML technologies and the privacy-preserving
characteristic of FL (due to its distributed learning nature) has resulted in increased demand
for FL with new technologies and frameworks currently being developed. For example,
Google has recently released a TensorFlow-based FL framework named TensorFlow Fed-
erated (TFF). TFF enables developers to deploy an AI system and train it across data from
multiple sources, all while keeping each of those sources separate and local (Ingerman and
Ostrowski, 2019). Other FL-based frameworks include Federated AI Technology Enabler
(Webank’s, AI, 2019), PySyft (Ryffel et al., 2018), Leaf (Caldas et al., 2019), PaddleFL
(Dong et al., 2019), and Clara Training Framework (Wen et al., 2019). In addition to the
4 MohammadNoor Injadat et al.
popular ML algorithms previous proposed in the literature such as neural networks and sup-
port vector machines, this illustrates the recent and continued development and deployment
efforts of ML algorithms and paradigms.
3 Education
The first area considered is that of the education sector. There are three main ways that
education can be delivered: onsite, online, and blended learning. Onsite education, or tradi-
tional education, refers to educational content delivery within a traditional classroom setting
(Moubayed et al., 2018). This setting requires that the educator and the students are in the
same room at the same time. This allows the educator to deliver his/her lecture to the at-
tending class. As such, traditional classrooms provide face-to-face interaction between the
educator and the students (Black, 2002).
On the other hand, online education, one category of e-learning systems, refers to edu-
cation that is provided over the Internet. E-learning provides students with the opportunity
to access educational curriculum outside of a traditional classroom at any time from any
geographical location. However, there is no face-to-face interaction with the educator as all
the content is delivered remotely (Moubayed et al., 2018).
Last but not least, blended or hybrid learning is a combination of onsite and online edu-
cation. For the education delivery system to be considered blended, up to 30% of the course
requirements must be conducted face-to-face in a traditional classroom setting, while the re-
maining percentage of the course requirements can be completed online. Blended learning
offers students the opportunity to have face-to-face interactions with the educator and other
students while also providing them with the opportunity to assess course materials at any
time from any location (Moubayed et al., 2018).
However, the educational sector faces a variety of challenges, some of which are ped-
agogical and others being technical. This section identifies some of the challenges in the
education sector. Moreover, it presents some of the previous literature works that tried to ad-
dress each of these challenges. Furthermore, it discusses the role of ML in addressing them.
challenges. More specifically, this section will discuss how ML can be used to grade es-
says, predict and prevent students from dropping-out, improve intelligent tutoring systems,
recommend online courses, and provide personalized learning.
3.1 Essay Grading
3.1.1 Challenge Description:
Essays provide a tool for assessing students’ critical thinking, analysis, and communica-
tion skills. However, it is time consuming for educators to grade essays. Furthermore, when
humans grade essays there is a great level of subjectivity which can lead to two different
graders scoring an essay very differently (Mahana et al., 2012). Within this context, using
ML algorithms to grade essays can reduce the workload of educators and provide more
objectivity during the grading process. A common approach to creating an essay grading
algorithm is to first collect a large pool of essays which have characteristics that are com-
putationally measurable (e.g., sentence length, word frequency distributions, grammar, and
spelling), and have been scored by humans (Ramalingam et al., 2018). This allows the algo-
rithm to first learn the characteristics which are important for grading an essay. Then, when
Machine Learning Towards Intelligent Systems: Applications, Challenges, and Opportunities 5
the algorithm is used to score essays, the algorithm’s scores can be compared to those of the
human graders in order to determine if the algorithm has properly learned the grading char-
acteristics. It is worth mentioning that this challenge is not only from the technical aspect,
but also from the social aspect in terms of accepting the output of the automated ML-based
models. However, this work only focuses on the technical aspect of this challenge rather
than the social aspect of it.
3.1.2 Previous Works:
Mahana et al. (2012) built an automated essay scoring system using essays from kaggle.com.
The authors selected roughly 13,000 essays from a pool of essays that were submitted to a
competition by the William and Flora Hewlett Foundation. The essays were written by stu-
dents from Grade 7 to Grade 10 and were approximately 150 to 550 words long. The selected
essays were divided into eight sets, with each of the sets having unique grading characteris-
tics. Eight different sets were selected to ensure that the automated grader was trained across
different types of essays. Furthermore, each essay has one or more human scores. In the lat-
ter case, the essay also had a final resolved score which considered all human scores. After
that, the authors selected the eight sets of training essays, they extracted several features
from them (e.g., total word count per essay, sentence count, number of long words, part of
speech counts, etc.) . These features were selected because they are characteristics that a
human grader would commonly look for when grading an essay. The authors then used a
linear regression model to allow their algorithm to learn parameters for grading based on the
selected features. After the algorithm learned the parameters for scoring the eight different
essay types, the algorithm was used to score a distinct set of test essays. These scores were
compared against human graded scores to arrive at an error metric (Quadratic Weighted
Kappa). The average kappa score of the authors’ algorithm across the eight essay types was
0.73. Essay set 8 had the lowest kappa at 0.68 and essay set 1 had the highest kappa at 0.80.
Despite the fact that the proposed framework achieved good performance as seen with the
high kappa values, this work has a limited contribution as it only considered one ML algo-
rithm.
Ramalingam et al. (2018) also used ML techniques to develop an automated essay as-
sessment system . The authors selected essays from a pool of essays that were submitted to
a competition by The Hewlett Foundation to kaggle.com. All of the essays had been graded
by humans. Similarly, the authors further segregated these essays into eight unique sets. This
was followed by using Bayesian Linear Regression as their algorithm. The Bayesian essay
scoring system used features like specific words, specific phrases, order in which certain
noun-verb pair appears, and the order of the concepts explained to score the essays. In the
end, the authors tested their algorithm on eighty essays divided into two groups of forty,
which had also been scored by humans. The authors’ algorithm was over 80% accurate at
scoring the essays. Yet, this work also only considered one ML technique which limits its
contribution.
Ullmann (2019) also discussed using ML to assess essays. More specifically, the author
investigated the potential of ML algorithms to automate the analysis of reflective essays.
To that end, the author explored eight different categories that are often used as metrics to
evaluate the quality of a reflective passage, namely reflection, experience, feeling, belief,
difficulty, perspective, learning, and intention. To that end, the author collected data from 76
students containing a total of 5080 sentences. Then, the authors created a training dataset
using a random sample consisting of 80% of the sentences and tested the performance of
four different ML classification algorithms including support vector machines (SVM), neu-
6 MohammadNoor Injadat et al.
ral networks, random forests (RF), and Naive Bayes on the remaining 20% of the sentences.
Experiments showed that the accuracy of the different ML models ranged between 80%-
96% for the different reflective essay metrics.
Mathias and Bhattacharyya (2020) explored the use of deep learning models to auto-
mate the essay grading process. To that end, the authors used the ASAP AEG dataset that
described differed essay sets with multiple essay traits such as Content, Organization, Word
Choice, Sentence Fluency, and Conventions (Mathias and Bhattacharyya, 2018). More-
over, the authors proposed the use of a feature engineering system, a string kernel, and
an attention-based neural network as part of their automated essay grading framework. Ad-
ditionally, the Cohen’s Kappa metric was used to evaluate the performance of the proposed
framework as it considers the random correct classification of data samples. Results showed
that the attention-based neural network algorithm outperformed other works from the lit-
erature across multiple prompts (with each prompt having a different set of essay traits)
with a Kappa value between 0.586 and 0.820. However, the work only considered deep neu-
ral networks without considering other potential classification algorithms that may be more
computationally efficient and have similar performance.
3.1.3 Research Opportunities:
As can be seen from this review of the literature, essays are an important tool for assessing
students’ comprehension and expression. However, grading essays is a time consuming task.
Furthermore, essay grading is prone to subjectivity, which can lead to the same essay being
scored differently by two graders. Hence, essay grading presents the challenges of time con-
sumption and human subjectivity.
ML offers a potential solution to address these issues. Firstly, ML algorithms can be
used so that graders no longer need to spend time on grading. Secondly, such algorithms can
be used to provide objective scores of essays. Although the previous subsection presented
research that has shown how ML can be used to address the challenges of essay grading,
there are still opportunities for further research.
One potential opportunity is exploring and evaluating different ML models (e.g. logistic
model trees or deep neural networks). This is mainly due to the fact that most of the previous
work only used one algorithm for essay grading. Therefore, it is important to explore and
compare the performance of other ML models to obtain a more robust essay grading frame-
work, especially given the effectiveness of other models such as deep neural networks in
natural language processing problems. Another potential opportunity is studying the impact
of more advanced Natural Language Processing features (e.g. N-grams, k-nearest neighbors
(k-NN) in bag of words), selecting features that are grammar and usage specific, and explor-
ing other polynomial basis functions like neural networks (NN) as part of the essay grading
framework.
Such frameworks can be applied to any assessment task that contains an essay compo-
nent. This includes exams and tests that contain essay sections. The application of essay
grading algorithms to exam and test essays could increase the consistency of scoring while
reducing the grader bias. Furthermore, there is the possibility to use essay grading algo-
rithms as components of interactive knowledge and writing tutorial systems.
Machine Learning Towards Intelligent Systems: Applications, Challenges, and Opportunities 7
3.2 Dropout Prevention
3.2.1 Challenge Description:
Student dropout is another challenge that is prevailing in the education sector. The term
dropout refers to the case when a student leaves/quits a course before completing it. Recent
studies showed that students were more likely to dropout of online courses than traditional
classes (Coussement et al., 2020). High dropout rates can effect the future of colleges and
universities. This is because different stakeholders within the education field including poli-
cymakers, funding bodies, and educators consider dropout rates to be an objective outcome-
based measure of the educational institutions’ quality (Sneyers and De Witte, 2017).
While the higher dropout rate of students in online classes is a known issue, there are
many possible reasons for students to dropout. In turn, this makes predicting dropout chal-
lenging. From the student side, possible reasons for online course dropout include higher
than expected workload, inability to manage academic responsibilities in a self-driven learn-
ing environment, unfamiliarity with the online educational delivery system, less student–teacher
interaction, family and social obligations, and motivation level (Bawa, 2016). However, stu-
dents may also dropout of online courses due to the course being poorly designed and de-
livered, which can occur when the professor who created and taught the online course is
unfamiliar with technology and/or is provided with no training by their institution on how
to teach in an online environment (Bawa, 2016).
As can be seen, there are many possible reasons students may dropout, which can make
predicting which students will dropout a complicated task. Even though this is a compli-
cated task, it is important to identify students at risk of dropping out so that professors can
address the needs of these students and take the appropriate actions to reduce their probabil-
ity of dropping out (Sneyers and De Witte, 2017). One way to make the task of identifying
at risk students easier is to use ML algorithms. Again, this challenge does not only entail the
technical aspect, but it also has a pedagogical aspect in terms of the context used for data
collection purposes. However, as mentioned earlier, this work only focuses on the technical
aspect with which ML can play a role without tackling the pedagogical context of the data
collection.
3.2.2 Previous Works:
Coussement et al. (2020) used ML techniques in order to predict dropout in e-learning
courses. More specifically, the authors proposed the use of logit leaf model (LLM), a deci-
sion tree (DT)-based classification model, to accurately predict student dropout in subscription-
based e-learning environments. LLM was chosen due to its capability to balance between
comprehensibility and predictive performance. To that end, the authors compared the per-
formance of the proposed LLM model to that of eight other ML classification algorithms
on a real-life dataset containing more than 10,000 students of a global subscription-based
e-learning provider. Results showed that the proposed LLM model was one of the top per-
forming student dropout prediction models with a high area under the curve (AUC) value
above 0.8, highlighting its effectiveness in achieving its target task.
Similarly, Chung and Lee (2019) proposed the use of an ML classification model to pre-
dict student dropout in high schools. In particular, they proposed the use of RF algorithm
for this task due to its high prediction accuracy in multiple scenarios and applications. To
that end, the authors used the National Education Information System (NEIS) data collected
in Korea in 2014 and evaluated the performance of the RF model using multiple metrics
8 MohammadNoor Injadat et al.
such as accuracy, sensitivity, specificity, and AUC. Experiments showed that the proposed
RF achieved high accuracy (close to 95%) and high AUC value (close to 0.97), highlighting
the effectiveness of this model. However, one shortcoming is that the work only considered
one classification algorithm without comparing it to other potential ML algorithms.
3.2.3 Research Opportunities:
Although ML has been proposed to predict dropout in e-learning courses, there are still
research opportunities within this area. One potential opportunity is studying the perfor-
mance of different ML dropout prediction frameworks and models in other course delivery
settings such as blended learning, distance and classical education. This would highlight
the generality of the dropout prediction framework. Another opportunity worth exploring
is investigating the impact of different student attributes to create their dropout prediction
method. This is essential as it can result in more accurate models. A third opportunity to
consider is comparing the performance of different base and ensemble learning methods to
achieve more accurate and robust prediction models and studying their impact on retention
strategies through correlation and association rules mining.
3.3 Intelligent Tutoring
3.3.1 Challenge Description:
The third challenge facing modern education systems is that of providing and improving
intelligent tutoring systems (Di Pietro and Distefano, 2019). For example, Troussas et al.
(2018) proposed an intelligent tutoring system to teach English and French languages using
ML-based models. However, in such a scenario, there are multiple challenges that arise
include how to detect spelling mistakes, verb tense mistakes, and auxiliary verb mistakes.
As such, developing effective intelligent tutoring systems can be a challenging task given
the significant impact of multiple factors that need to be considered.
3.3.2 Previous Works:
Di Pietro and Distefano (2019) combined the concepts of ML and gamification with cloud
technologies in a unified framework to improve intelligent tutoring systems. More specifi-
cally, the authors proposed the use of different ML models such as optical character recog-
nition, sentiment analysis, and speech recognition to create a virtual study buddy. The goal
of this system is to help students develop better study strategies by interacting with a digital
study partner. However, one limitation of this work was the fact that the authors did not test
their proposed framework to explore its effectiveness.
On the other hand, Barron-Estrada et al. (2017) proposed the use of sentiment analysis
to improve the performance of an affective intelligent tutoring system by better gauging the
opinions of students about the course contents. To that end, the authors used a collection
of texts containing more than 68,000 Twitter messages written in Spanish and transformed
them into numerical feature vectors along with their associated sentiment. Then, the authors
used Naive Bayes classifier (chosen due to its simplicity) to predict the sentiment of future
texts/Twitter messages. Again, the authors used multiple metrics such as accuracy, precision,
recall, and f1-score to evaluate the performance of the proposed module. Experimental re-
sults showed that their proposed module achieved high accuracy (above 80%) coupled with
Machine Learning Towards Intelligent Systems: Applications, Challenges, and Opportunities 9
high f1-scores. However, one shortcoming of this work is the fact that they did not compare
the performance of their proposed module to other potential classifiers.
3.3.3 Research Opportunities:
Despite the fact that ML has been used to improve intelligent tutors, more research opportu-
nities still exist. One such opportunity is considering more features (for example phonemic
and history-based features) and investigating their impact on the performance of the de-
veloped model. Another potential opportunity to consider is comparing the performance of
other classifiers such as NN and SVM. This comparison can help determine whether the
classifiers previously proposed in the literature are biased. Moreover, such comparisons will
lead to having a more adaptive and robust intelligent tutors.
3.4 Course Recommendation
3.4.1 Challenge Description:
Massively Open Online Courses (MOOCs) such us Coursera, Udacity, EdX, and MOOC.org
are a form of online distance education/e-learning. As such, MOOCs provide online courses
that can be accessed by a student at any time from any geographical location (Moubayed
et al., 2018). MOOCs are open to anyone that is interested in enrolling and are often free
or low-cost. However, the courses do not provide course credit and are not applied towards
a degree. Instead, MOOCs tend to be used by people who want to learn new skills, be it to
advance their career or for fun. MOOCs provide open access to a plethora of courses from
various top-rated universities and institutions. For example, the website mooc.org provides
a course titled “Data Science: R Basics” from Harvard University, and a course titled “In-
troduction to Data Analysis using Excel” from Microsoft (Mooc.org, 2019).
Since MOOCs are open access, hundreds of thousands of students can be enrolled in
each course, with MOOCs platforms offering thousands of different courses. This means
that MOOC platforms are privy to mass amounts of data. This data can then be used to im-
prove the MOOC system. For example, having thousands of different courses available can
be overwhelming for students. Therefore, if a student is looking to improve a specific skill,
it would be beneficial for the MOOC system to recommend which courses are needed to
acquire those skills (Symeonidis and Malakoudis, 2016).
3.4.2 Previous Works:
Several previous literature works focused on the problem of course recommendation for
students. One such example is the work by (Aher and Lobo, 2013). The authors used prior
student data and a combination of ML algorithms to recommend courses to students in
an e-learning system. The authors combined Simple K-means (a clustering technique) and
Apriori (an association rule algorithm) to investigate prior students’ data from Moodle.org in
order to determine which courses to recommend to new students. The authors found that the
results of their combination approach matched real world student course selection patterns.
However, one limitation of this work is that it only considered one unsupervised clustering
algorithm.
Mondal et al. (2020) also proposed the use of ML algorithms as part of a course rec-
ommendation system for online learning environments. More specifically, the authors first
10 MohammadNoor Injadat et al.
used K-means algorithm to group students based on their performance in previous courses.
This was followed by applying collaborative filtering to recommend new suitable courses.
The results showed that the proposed model achieved a low root mean squared error and
mean absolute error. Furthermore, the model also achieved high precision and recall values,
indicating that it can return correct results and preserve the majority of true positives.
Similarly, Zhang et al. (2018a) proposed the use of a distributed association rule algo-
rithm as part of their course recommendation system. The authors used a combination of
Hadoop and Spark platforms to implement the proposed framework so that it is suitable for
MOOC environments. The experimental results on three different datasets illustrated the ef-
fectiveness of the proposed framework by having a high confidence value (close to 0.5) for
multiple association rules.
3.4.3 Research Opportunities:
ML algorithms can be further applied to the large amounts of data that MOOC platforms
possess in order to determine which courses would be best for a student who is interested
in improving a specific skill set. One potential research opportunity for students’ course
recommendation is evaluating the courses that other students have taken that are related to
the skill that the student is interested in. Using that information can help build an effective
course recommender. Another opportunity is consider multiple supervised classification al-
gorithms. This is particularly important given the substantial impact that the classification
process has on the overall performance of the recommender. Therefore, it is worth exploring
the performance of different classification algorithms to study their impact on the effective-
ness of the recommendation process.
3.5 Personalized Learning
3.5.1 Challenge Description:
Personalized learning is based on the individual students and how they learn. Each individual
learns differently and has a unique learner profile. This profile is based on the individual’s
learning style (Klaˇ
snja-Mili´
cevi´
c et al., 2011; Dwivedi and Bharadwaj, 2015; Bourkoukou
and El Bachari, 2016), which consists of specific behaviors and attitudes (Truong, 2016).
Personalizing each learner’s education can lead to better learning. One way to personal-
ize education is by using recommender systems that provide useful suggestions for users
(books, movies, products, etc.) based on their preferences and their similarity to other stu-
dents (Bourkoukou and El Bachari, 2018).
3.5.2 Previous Works:
Bourkoukou and El Bachari (2018) tested the ability of LearnFitII to act as a recommender
system. LearnFitII is an adaptive learning system that automatically adapts to the dynamic
preferences of learners. By mining the server logs of students, LearnFitII was able to rec-
ognize the different learning styles and habits of students. Then, using the Felder-Silverman
model of learning styles, LearnFitII proposed personalized learning scenarios. The Felder-
Silverman model of learning styles consists of four learning dimensions (1. Information
Processing, 2. Information Perception, 3. Information Reception, and 4. Information Under-
standing) (Felder et al., 1988). These dimensions can be accessed via the Index Learning
Machine Learning Towards Intelligent Systems: Applications, Challenges, and Opportunities 11
Style Questionnaire (ILSQ) which consists of 44 questions.
After proposing personalized learning scenarios, LearnFitII analyzed the habits and the
preferences of learners by mining information about the learners’ actions and interactions.
After the mining of this information, the learning scenarios were revisited and updated us-
ing a hybrid recommender system which combined k-NN and association rule mining algo-
rithms. The authors found that when LearnFitII was tested in real environments that learning
quality increased and so did the learners’ satisfaction with the learning process (Bourkoukou
and El Bachari, 2018).
Another way that personalized learning can be beneficial is in helping students select the
learning-pathway that is appropriate for them. Elfaki et al. (2014) investigated how student
learning-pathways can be improved with ML. The term learning-pathway can be under-
stood as the path of academic courses that is appropriate for a student to achieve a degree.
Ideally, one’s learning-pathway is in their field of interest. Typically, students spend some
time taking various courses in order to discover which topics they are interested in. How-
ever, this process of taking various courses can lead to a mismatch between a student’s
current and preferred learning pathway. When mismatches occur, the student may experi-
ence academic difficulties (e.g. weak performance, high absentee rate). These mismatches
may lead students to lower their level of education or dropout of university altogether. In
order to improve students’ levels of achievement it would be beneficial to help them deter-
mine their desired learning-pathway sooner. In order to achieve this goal sooner, the authors
first collected questionnaire data. Then they sent a questionnaire to 900 students from the
Faculty of Computers and Information Technology at Tabuk University in Saudi Arabia with
450 students returning the questionnaire. The questionnaire addressed four topics: basic in-
formation, personal information, academic information, and learning pathway information.
After collecting this data, the authors applied a DT algorithm to the data. Then, induction
rules were deduced from the tree paths in order to provide learning-pathway recommen-
dations. In order to validate their results, the authors divided the questionnaires into two
groups, a developing group (70%) and a test group (30%). Using these two groups of data,
the authors found that their algorithm could accurately provide learning-pathway recom-
mendations (Elfaki et al., 2014).
Moubayed et al. (2018); Moubayed et al. (2019) studied the problem of student en-
gagement level identification in an e-learning environment. This was done using K-means
algorithm. In addition to that, the authors extracted a set of rules relating student engage-
ment to academic performance. To that end, the authors used the Apriori association rules
algorithm. Experimental results showed a positive relationship between students’ engage-
ment level and their academic performance.
Injadat et al. (2020c,a) proposed the use of optimized ML ensemble classification mod-
els to predict student performance during the course delivery time at two stages. The au-
thors explored the use of various based learners such as SVM, K-NN, NB, RF, and neural
networks to form the ensembles and tested them on two different datasets. Results showed
that their proposed ensemble models achieved high accuracy for the target class in both the
binary and multi-class cases despite the small number of instances available.
3.5.3 Research Opportunities:
There are still further research opportunities to use ML to provide personalized learning.
One opportunity is to consider more complex recommendation approaches by including
other factors such as learner motivation and knowledge level as well as additional personality
traits. Another opportunity is to study the performance of different classification algorithms
12 MohammadNoor Injadat et al.
Table 1: Challenges, Previous Works, and Research Opportunities within Education Sector
Challenge Previous Work Research Opportunity
Essay Grading
Regular linear regression is used for
essay grading (Mahana et al., 2012)
- Explore different ML models such
as LR, DT, and DNN
Bayesian linear regression is used
for essay grading (Ramalingam
et al., 2018)
- Study the impact of additional
Language and usage-specific fea-
tures as well as other polynomial
basis functions.
SVM, RF, neural networks, and
naive bayes are used to evaluate re-
flective essays (Ullmann, 2019)
- Compare the performance of the
different models on various tasks to
get a more accurate and robust es-
say grading framework.
Deep neural networks were used to
classify essays based on five dif-
ferent potential traits (Mathias and
Bhattacharyya, 2020)
Dropout Prevention
Studied the performance LLM to
predict dropout (Coussement et al.,
2020)
- Study the performance of different
models (base learners and ensemble
learner models) in different course
delivery settings.
Studied the performance of RF clas-
sification model for dropout predic-
tion(Chung and Lee, 2019)
- Investigate the impact of differ-
ent student attributes on the dropout
prediction frameworks.
Intelligent Tutors Sentiment analysis was proposed
as part of a virtual study buddy
framework (Di Pietro and Diste-
fano, 2019)
- Consider more features such as
phonemic and history-based fea-
tures as well as investigate their im-
pact on the performance of the de-
veloped model.
NB was used to predict the sen-
timent of text messages to be
used to improve an affective in-
telligent tutoring system (Barron-
Estrada et al., 2017)
- Study the performance of other
classifiers such as NN and SVM
to determine whether the classifiers
previously proposed in the litera-
ture are biased.
Course
Recommendation
Combined k-means and apriori al-
gorithm to recommend courses
(Aher and Lobo, 2013)
- Evaluate the courses that other stu-
dents have taken that are related to
the skill the student is interested in
using multiple metrics.
Used a combination of K-means al-
gorithm and collaborative filtering
for course recommendation (Mon-
dal et al., 2020)
- Consider multiple supervised clas-
sification algorithms to study their
impact on the effectiveness of rec-
ommendation process.
Used an improved version of
Apriori association rules algorithm
for course recommendation (Zhang
et al., 2018a)
Personalized
Learning
Combined a DT algorithm and in-
duction rules algorithm to pro-
vide learning-pathway recommen-
dations (Elfaki et al., 2014)
- Study the performance of differ-
ent classification algorithms to pre-
dict student performance during the
course delivery.
Mined server logs to determine stu-
dent learning style (Bourkoukou
and El Bachari, 2018)
- Consider more complex recom-
mendation approaches by including
other factors.
Used K-means and apriori al-
gorithms to identify student en-
gagement and their relation with
academic performance (Moubayed
et al., 2018; Moubayed et al., 2019)
Used multiple optimized ensemble
classification algorithms to predict
student performance during course
delivery (Injadat et al., 2020c,a)
Machine Learning Towards Intelligent Systems: Applications, Challenges, and Opportunities 13
Fig. 2: Potential Deployment of ML in Learning Management System (LMS)
to predict student performance during the course delivery. This can help identify student
who may need help and provide them with a personalized plan to improve their predicted
performance.
Table 1 summarizes the challenges within the education sector, lists some of the previous
works, and presents the different research opportunities. Furthermore, Figure 2 illustrates
the potential deployment framework of the ML modules within the Learning Management
System (LMS).
4 Healthcare
Another area where ML has shown promise is in the field of healthcare. Many modern
medical organizations use electronic health records (EHRs) (Caban and Gotz, 2015), EHRs
consist of heterogeneous data elements. This includes information such as the patient demo-
graphic, diagnoses, laboratory test results, previous prescriptions, and clinical notes (Xiao
et al., 2018). Patient data can also include imaging, sensor and text data (Miotto et al., 2018).
Furthermore, this data often comes in various formats, including structured, semi-structured
and weakly structured data (Holzinger et al., 2014). Originally it was thought that having
access to more information about individual patients would lead to more informed medi-
cal decisions. However, often times health professionals are overwhelmed by the amount
of information that is now available to them (Caban and Gotz, 2015). Hence, a challenge
with big data in healthcare is making the data easily interpretable for medical professionals.
ML offers a solution to this problem because it can be used to identify relevant patterns in
complex data. In this section, how ML algorithms can be used in various applications such
as predicting individual patient’s responses to cancer drugs, diabetes research, retinopathy
detection, and cancer detection is discussed. Note that despite the various applications in
which ML can be applied within the field of healthcare, this section discusses some of the
most prominent applications.
14 MohammadNoor Injadat et al.
4.1 Drug Response Prediction
4.1.1 Challenge Description:
One way that ML can be applied to medical data is to predict an individual patient’s response
to a drug or drugs (Vidyasagar, 2015). For example, ML can be used to predict the responses
of individual cancer patient to therapeutic drugs (Huang et al., 2018). When working with
cancer patients, it is possible to use precision cancer medicine. Precision cancer medicine
aims to accurately predict the optimal drug therapies for a patient based upon the person-
alized molecular profiles of their tumors (Prasad et al., 2016). In order to provide precision
cancer medicine, it is necessary to search for significant correlations between patient tumor
profiles and the output predictions of optimal drug responses in cancer-relevant datasets.
Once these correlations are found in previously established datasets, they can be used to
predict an individual patient’s response to various series of therapeutic drugs (Vidyasagar,
2015). As mentioned before, ML offers a solution to this problem, because it can be used to
identify relevant patterns in complex data.
4.1.2 Previous Works:
Huang et al. (2018) applied their open-source SVM-based algorithm to the gene-expression
profiles of 175 individual cancer patient’s tumors. The algorithm was able to predict the
responses of these 175 individuals to a variety of standard-of-care chemotherapeutic drugs
with >80% accuracy.
Xia et al. (2018) also used ML to predict tumor cell line response to drug pairs. The
authors used a computational deep learning model to predict cell line response to a subset of
drug pairs in the National Cancer Institute-ALMANAC database. When the authors ranked
the drug pairs for each cell line based on the model’s predicted combination effect, they
were able to determine 80% of the top drug pairs.
Chiu et al. (2019) also used deep neural networks (DNN) and the genomic profiles of
cancer tumors in order to predict the tumors’ responses to therapeutic drugs. The authors
created DeepDR, a deep neural network model, then trained it to learn the genetic back-
ground of tumors based on data from The Cancer Genome Atlas (TCGA). DeepDR was also
trained on pharmacogenomics data from human cancer cell lines provided by the Genomics
of Drug Sensitivity in Cancer (GDSC) Project. After training on these data sets, DeepDR
was applied to TCGA data again in order to predict the drug response of tumors. The au-
thors’ work provides insights into the ability of a deep neural network model to translate
pharmacogenomics features identified from in vitro drug screening to predict the response
of tumors.
Mucaki et al. (2019) used ML and genetic data to predict patients’ responses to chemother-
apy. More specifically, the authors used supervised support vector ML to determine the gene
sets whose expression was related to the specific tumor cell line GI50 . The authors discov-
ered that specific genes and functional pathways can be used to distinguish which tumor
cell lines are sensitive to chemotherapy drugs and which tumor cell lines are resistant to
chemotherapy drugs. They tested their algorithm on bladder, ovarian and colorectal cancer
patient data from The Cancer Genome Atlas (TCGA) in order to determine the response
of tumor cell line GI50 to three chemotherapy drugs (cisplatin, carboplatin and oxaliplatin).
Through experimental results, the authors found that for cisplatin, their algorithm was 71.0%
accurate at predicting disease recurrence and 59% accurate at predicting remission. In the
case of carboplatin, their algorithm was 60.2% accurate at predicting disease recurrence and
Machine Learning Towards Intelligent Systems: Applications, Challenges, and Opportunities 15
61% accurate at predicting remission. Finally, for oxaliplatin, their algorithm was 54.5%
accurate at predicting disease recurrence and 72% accurate at predicting remission. Fur-
thermore, in patients who used cisplatin and had a specific genetic signature, the algorithm
was able to predict 100% of recurrence in non-smoking bladder cancer patients and 79%
recurrence in smokers.
4.1.3 Research Opportunities:
Many research opportunities still exist in applying ML for drug response prediction. One
such opportunity is extending existing models to to predict the drug responses of cancer
patients who are receiving emerging immuno- and other targeted gene therapies. This will
validate the comprehensiveness and generality of the considered frameworks. Another po-
tential research opportunity is to build more comprehensive models by using more drug
features (such as concentration, SMILES strings, molecular graph convolution and atomic
convolution). This again will help extract more information and potentially uncover more
correlations and inter-dependencies that can make the models more robust and accurate. A
third opportunity is to investigate other methods and techniques including semi-supervised
learning methods to encode molecular features with external gene expression and other types
of data. This particularly would be helpful given that access to labeled data is not always
possible. Therefore, having semi-supervised based ML models can help healthcare profes-
sionals gain insight from labeled data and apply it to the unlabeled data that they have.
Last but not least, researchers should also investigate ways to adapt existing models to other
drugs, cancer types, and diseases. This is essential as it would provide one adaptive sys-
tem that can help healthcare professionals from different specializations make use of the
available data.
4.2 Diabetes Research
4.2.1 Challenge Description:
As mentioned before, many modern medical organizations use electronic health records
(EHRs) to store the medical data of patients. The large amounts of data present in EHRs
can be a valuable source for researching diabetes mellitus (DM). Kavakiotis et al. (2017)
discuss what DM is and why it is a medical concern. DM is a group of metabolic disorders
that are mainly caused by abnormal insulin secretion and/or action. Abnormal insulin secre-
tion can result in a patient’s body not producing enough insulin which causes the patient’s
metabolism of carbohydrates, fat and proteins to be impaired, which in turn results in ele-
vated blood glucose levels (hyperglycaemia).
There are two major clinical types of DM, type 1 diabetes (T1D) and type 2 diabetes
(T2D). T1D is linked to the auto-immunological destruction of the Langerhans islets; whereas
T2D is linked to lifestyle, little physical activity, poor dietary habits and heredity. The main
treatment for T1D is insulin administration which can applied to T2D patients. However,
the main treatment for T2D is improved diet, weight loss, exercise and oral medication.
DM affects more than 200 million people worldwide, with 10% of those affected with T1D
and 90% affected with T2D. DM possess a health threat as chronic hyperglycaemia results
in several complications, including diabetic nephropathy, retinopathy, neuropathy, diabetic
coma and cardiovascular disease.
16 MohammadNoor Injadat et al.
4.2.2 Previous Works:
DM has a high mortality and morbidity rate, therefore, detecting and treating DM is of high
interest to the medical community as well as those who may or already do suffer from DM
(Kavakiotis et al., 2017). In recent years, researchers have been able to apply ML algorithms
to the data of patients with DM in order to improve the methods of detecting and treating
DM.
Hemoglobin is a substance in red blood cells that carries oxygen to tissues. However,
it can also attach to sugar in the blood and form a substance called glycated hemoglobin
(HbA1c) (Michael Dansinger, 2019). A patient’s HbA1c level can be checked in order to
determine if they have T2D. Alternatively, a patient’s HbA1c level along with their fasting
blood glucose level and oral glucose tolerance test results can be used to determine if they
have T2D (Jelinek et al., 2016). Currently, to diagnosis a patient with T2D their HbA1c
value must be at or above 6.5% . However, studies have shown that the cut-offvalue of 6.5%
leads to inconsistencies in the diagnosis of T2D. Hence, using HbA1c with a 6.5% cut-off
value as a single marker for T2D may lead to undiagnosed cases of diabetes.
Jelinek et al. (2016) applied ML algorithms to the data of 840 patients from the Diabetes
Health screening (DiabHealth) in order to identify an optimal cut-offvalue for HbA1c and
to identify whether additional biomarkers could be used along with HbA1c to increase the
diagnosis of T2D. Then the authors used T2D as the class feature and generated a conven-
tional DT using an information gain (IG) measure. Using this algorithm, the authors found
that if an oxidative stress marker (8-OhdG) was included in the model along with HbA1c that
the accuracy of detecting T2D at the 6.5% HbA1c level increased from 78.71% to 86.64%.
The authors also found that if interleukin-6 (IL-6) was included in the model along with
HbA1c that the accuracy of detecting T2D increased from 78.71% to 85.63%. However, in
this model, the optimal HbA1c range was between 5.73 and 6.22% .
Herrero et al. (2014) used ML to improve the treatment methods of T1D. Diabetics
with T1D need to use the medication insulin in order to maintain normal blood sugar lev-
els. Diabetics must self-administer multiple daily injections of insulin, both before meals
and basally, in order to mimic the natural insulin secretion of the pancreas. Before diabet-
ics administer these injections, they must prick their fingertip to draw blood that is placed
in an electronic glucose meter that determines the amount of glucose in the patient’s blood
(Schiffrin and Belmonte, 1982; A. Flyvbjerg and Goldstein, 2010). However, in recent years,
an alternative form of therapy has become available. This alternative form of therapy is in-
sulin pump therapy. In this therapy, injections are provided by continuous subcutaneous
insulin infusion. The application of insulin by a machine allows diabetics to avoid multiple
uncomfortable finger pricks and injections. Insulin that is taken at meal times is referred to as
bolus insulin. Typically, bolus insulin doses are calculated by estimating carbohydrate intake
and dividing this number by a fixed carbohydrate to insulin ratio, then adding a correction
dose derived from the individual’s insulin sensitivity factor (Herrero et al., 2014). Although
several algorithms have been developed to calculate bolus insulin dose (Jovanovic and Pe-
terson, 1982; Chanoch et al., 1985; Schiffrin et al., 1985; Chiarelli et al., 1990; Albisser,
2003; Owens et al., 2006), these algorithms have only been incorporated in commercially
available insulin pumps and in some glucose meters (Zisser et al., 2008). However, these al-
gorithms have not been adopted widely commercially due to economic risk, security issues
and inertia to change, and the lack of ease of use (Bellazzi, 2008). Based on these challenges,
Herrero et al. (2014) set out to create a more user-friendly bolus insulin calculating system.
The authors used a decision support algorithm that incorporated Run-To-Run (R2R) control
and case-based reasoning (CBR). They tested their algorithm via in-silico scenarios by using
Machine Learning Towards Intelligent Systems: Applications, Challenges, and Opportunities 17
a simulator that emulated intra-subject insulin sensitivity variations and uncertainty in the
capillarity measurements and carbohydrate intake. Via these simulations, the authors found
that the CBR(R2R) algorithm significantly reduced the mean blood glucose level and com-
pletely eliminated hypoglycemia. When the authors compared the CBR(R2R) algorithm to
a standalone (R2R only) version of the algorithm, they found that the CBR(R2R) algorithm
performed better. The goal of the algorithm was to reduce blood glucose levels. Therefore,
the CBR(R2R) algorithm performed better than the standalone R2R algorithm in both pop-
ulations.
Fig. 3: Potential Deployment of ML in Diabetes Research
4.2.3 Research Opportunities:
Despite the fact that ML has been used in diabetes research, more opportunities still exist.
One suggestion is testing existing algorithms in the real-world via clinical trials. This is
particularly important given that simulation environment tend to over-estimate the benefits
of an intervention and may not always provide an accurate representation of the behavior of
the body. Another opportunity worth exploring is investigating the performance of different
ML classification models. This can help validate whether existing models have any bias.
Therefore, it is important to compare the performance of different models to have a more
accurate and sensitive model for insulin calculation. Figure 3 provides a visualization of
how these topics fit into a precision medicine framework.
18 MohammadNoor Injadat et al.
4.3 Retinopathy Detection Through Image Classification
4.3.1 Challenge Description:
Another healthcare-related area in which ML is playing a major role is for the detection of
retinopathy. Retinopathy refers to the damage of the retina, the light-sensing inner part of the
human eye (Harvard Medical School, 2017). This can be due to multiple causes and diseases
which can lead to partial or complete vision loss. There are different types of retinopathy
including: retinopathy of prematurity, diabetic retinopathy, hypertensive retinopathy, and
central serous retinopathy. Detecting the damage at an early stage is crucial to facilitate
the treatment and slow down the loss vision process. To that end, multiple previous works
proposed the use of ML algorithms and paradigms to accurately detect retinopathy.
4.3.2 Previous Works:
Bhatia et al. (2016) proposed the use of ensemble ML models to perform early detection of
diabetic retinopathy. More specifically, the authors proposed ensembles based on DT, ad-
aBoost, Naive Bayes, K-NN, RF, and SVM to detect this disease using features extracted
from retinal images such as diameter of optic disk, lesion specific, and image level features.
Their experiments showed that the proposed models achieved a detection accuracy of up to
94%.
Similarly, M¨
uller et al. (2020) also proposed the use of ensemble models to detect
ABCA4-Related Retinopathy. The authors used high-dimensional microstructural eye im-
age’s dataset and extracted multiple features. The authors then developed different ensemble
learning models based on K-NN, RF, SVM, and eXtreme Gradient Boosting (XGBoost).
Their experimental results showed that the proposed model achieved detection accuracies
ranging between 86%-93%.
Reddy et al. (2020) also proposed the use of ensemble-based ML models to detect di-
abetic retinopathy. In their work, the authors considered multiple classifiers including RF,
DT, Adaboost, K-NN, and Logistic Regression (LR). These methods were applied to a di-
abetic retinopathy dataset that was normalized using the min-max method. Results showed
that the proposed ensemble model outperformed the base models and achieved a detection
accuracy above 80%.
On the other hand, Gadekallu et al. (2020) proposed the combination of principal com-
ponent analysis (PCA) and DNN for the early detection of diabetic retinopathy. The authors
used a diabetes retinopathy dataset available at the UCI machine learning repository and nor-
malized it using the Z-score technique. After normalization, PCA was applied to extract the
most significant features. The reduced dataset was then given as an input to a DNN model
for classification. The experimental results showed that the proposed DNN model achieved
training accuracies between 72%-82% and testing accuracies between 68%-79%.
Similarly, Zhang et al. (2019) used DNN for the automated identification and grading
system of diabetic retinopathy. The proposed system uses transfer learning and ensemble
learning to detect the presence and severity of DR from fundus images. The authors’ ex-
perimental results showed that their developed model has a high identification sensitivity of
97.5% and a specificity of 97.7%. On the other hand, the grading model achieved a sensitiv-
ity of 98.1% and a specificity of 98.9%.
Machine Learning Towards Intelligent Systems: Applications, Challenges, and Opportunities 19
4.3.3 Research Opportunities:
Despite the promising results shown by ML models for retinopathy detection, further op-
portunities still exist. One such opportunity is developing optimized ML models. This is
because many of the previous works in the literature only consider the default version of the
classifiers. However, it is important to explore the impact of different optimization methods
on the overall performance of the classifiers. Another opportunity is to consider different
deep learning models such as convolutional neural networks (CNN) and recurrent neural
networks (RNN). More specifically, CNNs and RNNs can be combined to extend the ca-
pabilities of traditional CNN models from the binary to multi-label image classification as
illustrated in (Wang et al., 2016; Shi and Pun, 2018). Such architectures have the potential to
further improve the performance of deep learning models for early detection of retinopathy.
4.4 Cancer Detection Through Image Classification
4.4.1 Challenge Description:
Another prominent area in which ML is used within the healthcare field is the detection of
different types of cancer including breast cancer, prostate cancer, and lung cancer. This is
often done by applying ML methods to images of the different organs or tissue suspected to
have cancer (Saba, 2020). Due to their success in image classification problems in general,
ML models have been proposed in multiple research works from the literature to detect
cancer based on tissue images.
4.4.2 Previous Works:
Agarap (2018) investigated the performance of six different ML algorithms to detect breast
cancer. More specifically, the author compared between linear regression, multi-layer per-
ceptron (MLP), K-NN, softmax regression, and two variants of SVM algorithm. To evaluate
the performance of these algorithms, the Wisconsin diagnostic breast cancer dataset was
used which is composed of features extracted from digitized images of tests on breast mass.
Experimental results showed that the detection accuracy ranged between 93%-99% with the
MLP algorithm being the most accurate.
Similarly, Shen et al. (2019) also proposed the use of ML algorithms for breast can-
cer detection. However, the authors in this case developed a deep learning model, namely
CNN, that was applied to digitized film mammograms from the Digital Database for Screen-
ing Mammography. Experimental results showed that the developed CNN model achieved
accuracies between 63%-99% with the performance dependent on whether the CNN was
pre-trained or not. Moreover, the developed model achieved an area under the curve (AUC)
value reaching 0.88, illustrating its effectiveness and robustness in detecting breast cancer.
On the other hand, Hussain et al. (2018) proposed the use of ML models for prostate can-
cer. To that end, the authors explored different ML algorithms such as Bayesian approach,
SVM with multiple kernels, and DT. The authors also investigated different feature extrac-
tion strategies based on texture, morphological, scale invariant feature transform (SIFT),
and elliptic Fourier descriptors (EFDs) features. Experimental results showed that the SVM
classifier with RBF kernel achieved the highest accuracy ranging between 98%-99%.
Wu and Zhao (2017) proposed the use of ML algorithms to detect lung cancer based on
20 MohammadNoor Injadat et al.
computed tomography (CT) scan images. To that end, the authors proposed the use of a neu-
ral network-based model, namely the entropy degradation method (EDM), to detect cancer-
ous images. The performance of the proposed model was explored using a high-resolution
CT scan images provided by the National Cancer institute. Experimental results showed that
the proposed model achieved a detection accuracy of 77.8%, illustrating its effectiveness to
detect small-cell lung cancer at an early stage.
Table 2: Challenges, Previous Works, and Research Opportunities within Healthcare Sector
Challenge Previous Work Research Opportunity
Drug
Response
Prediction
Applied (SVM)-based algorithm to pre-
dict the responses of individuals to a
variety of standard-of-care chemothera-
peutic drugs (Huang et al., 2018)
- Extend existing models to to predict
the drug responses of cancer patients
who are receiving emerging immuno-
and other targeted gene therapies.
Developed a deep learning model to pre-
dict cell line response to a subset of drug
pairs in the National Cancer Institute-
ALMANAC database (Xia et al., 2018)
- Build more comprehensive models by
using more drug features.
Used DNN and the genomic profiles of
cancer tumors to predict the tumors’ re-
sponses to therapeutic drugs (Chiu et al.,
2019)
- Investigate other methods and tech-
niques including semi-supervised learn-
ing methods on other types of data.
Used SVM to determine the gene sets
whose expression was related to the spe-
cific tumor cell line GI50 (Mucaki et al.,
2019)
- Investigate ways to adapt existing
models to other drugs, cancer types, and
diseases.
Diabetes
Research
Used conventional DT to identify an op-
timal cut-offvalue for HbA1c (Jelinek
et al., 2016)
- Investigate the performance of differ-
ent ML classification models to validate
whether existing models have any bias.
Used a decision support algorithm to
calculate the bolus insulin levels (Her-
rero et al., 2014)
- Test existing algorithms in the real-
world via clinical trials.
Retinopathy
Detection
Through
Image
Classification
Used of ensemble ML models based
on multiple algorithms to perform early
detection of diabetic retinopathy(Bhatia
et al., 2016)
- Develop optimized ML models to fur-
ther improve detection accuracy.
Used ensemble models to detect
ABCA4-Related Retinopathy (M¨
uller
et al., 2020)
- Consider different deep learning mod-
els and architectures such as CNN and
RNN.
Proposed ensemble-based ML models
to detect diabetic retinopathy(Reddy
et al., 2020)
- Use DNN techniques to build models
that predict if a patient is diabetic or not.
Proposed the combination of PCA and
DNN for the early detection of diabetic
retinopathy (Gadekallu et al., 2020)
Used DNN for the automated identifi-
cation and grading system of diabetic
retinopathy (Zhang et al., 2019)
Cancer
Detection
Through
Image
Classification
Investigated the performance of six dif-
ferent ML algorithms to detect breast
cancer (Agarap, 2018)
- Develop optimized ML models to fur-
ther improve detection accuracy.
Proposed a deep learning CNN model
for breast cancer detection (Shen et al.,
2019)
- Apply different types of ensemble
models that can combine multiple ML
classifiers to improve their effectiveness
and robustness.
Explored different ML classifiers to de-
tect prostate cancer (Hussain et al.,
2018)
- Consider different deep learning mod-
els and architectures such as RNN.
Proposed the EDM (a neural network-
based model) model to detect small-cell
lung cancer (Wu and Zhao, 2017)
- Explore different algorithms to im-
prove the feature extraction and selec-
tion process.
Machine Learning Towards Intelligent Systems: Applications, Challenges, and Opportunities 21
4.4.3 Research Opportunities:
As shown above, ML algorithms have been successfully applied to detect different types
of cancer. However, there still exists further opportunities to improve the detection perfor-
mance. One such opportunity is to consider different hyper-parameter optimization methods
to improve the performance of the ML models. Another potential opportunity is applying
different types of ensemble models that can combine multiple ML classifiers to improve their
effectiveness and robustness. A third opportunity is studying other deep learning techniques
and architectures such as the combined CNN-RNN models to investigate their effectiveness
in detecting cancer. An additional opportunity is to improve the feature extraction and se-
lection algorithms from digital images. This is crucial given the fact that this process acts as
the input to the ML model development stage. Therefore, it is important to extract and select
relevant and high-quality features to be fed to the ML models under consideration.
Similar to the previous section on education, Table 2 summarizes some of the challenges
facing the healthcare sector, lists some of the previous works, and presents the different re-
search opportunities.
5 Network Security
Turning to a different sector, ML can also be beneficial in network security. Cisco Systems,
Inc., an American multinational technology conglomerate who specializes in information
technology, networking, and cybersecurity solutions, defines network security as any ac-
tivity designed to protect the usability and integrity of a network and data (Cisco, 2019).
According to Cisco Systems, Inc., network security allows authorized users to assess a net-
work while preventing outside threats from entering or spreading on a network (Moubayed
et al., 2018, 2020). Cisco Systems, Inc. lists fourteen types of network security. However,
this section will focus on Intrusion Detection Systems (IDS). IDSs analyze and monitor net-
work traffic in order to determine if the network traffic patterns show normal activity or if
there are signs of malicious activity (Javaid et al., 2016; Sommer and Paxson, 2010). More
specifically, this section will discuss how ML can be used to improve network intrusion de-
tection systems (NIDS) in general, how to better detect Botnets, and how to improve NIDS
in vehicles.
5.1 Network Intrusion Detection Systems
5.1.1 Challenge Description:
A Network Intrusion Detection System (NIDS) helps system administrators to detect net-
work security breaches in their organizations (Javaid et al., 2016; Salo et al., 2018). NIDSs
are classified based on the style of detection that they use. Misuse-detection NIDSs use pre-
cise descriptions of known malicious behavior. Anomaly-detection NIDSs flag deviations
from normal activity. Specification-based NIDSs define allowed types of activity and flag
any other activity as forbidden. Behavioral detection NIDSs analyze patterns of activity and
surrounding context to find secondary evidence of attacks. Although there are many types
of NIDS, misuse-detection and anomaly-detection NIDSs are the most common (Sommer
and Paxson, 2010).
Misuse-detection NIDSs can also be referred to as signature (misuse) based NIDS (SNIDS).
22 MohammadNoor Injadat et al.
In SNIDS, attack signatures are pre-installed in the NIDS and pattern matching is then per-
formed between network traffic and the installed signatures. When a mismatch is found,
it is considered an intrusion. There are advantages and disadvantages to both SNIDS and
anomaly-detection NIDS (ADNIDS). SNIDSs are effective in the detection of known at-
tacks and show high detection accuracy with less false-alarm rates, but is ineffective at de-
tecting unknown or new attacks whose signatures have not been installed on the IDS. On the
other hand, ADNIDSs are the better option for the detection of unknown and new attacks,
but produce high false-positive rates. The current deployment framework and usage pat-
ters of NIDSs makes it hard for these systems to be efficient and flexible when considering
unknown future attacks (Javaid et al., 2016).
5.1.2 Previous Works:
Javaid et al. (2016) propose a solution to the challenges of using NIDS to detect known
and unknown future attacks. The authors’ solution involved using a deep learning approach
known as Self-Taught Learning (STL). The authors verified their method on the bench-
mark intrusion dataset NSL-KDD. This dataset is an improved version of the former bench-
mark intrusion dataset KDD Cup 99. The authors present various metrics related to their
algorithm, including accuracy, precision, recall, and f-measure values. Experimental results
showed that the authors’ algorithm achieved a classification accuracy rate above 98%.
Injadat et al. (2018) proposed using Bayesian optimization to hyper-tune the parame-
ters of different supervised ML algorithms for anomaly-based IDSs. More specifically, they
tune the parameters of SVM, Random Forest (RF), and k-NN algorithms. Then, the authors
evaluated the performance of the regular and optimized version of these classifiers in terms
of accuracy, precision, and false alarm rate. Their experimental results showed that the pro-
posed framework achieved a high accuracy rate and precision, and a low-false alarm rate
and recall.
Injadat et al. (2020b) extended the work by proposing a novel multi-stage optimized ML-
based NIDS framework. The goal of this framework is to reduce the computational complex-
ity and maintain the detection performance. The performance of the proposed framework
was measured using two state-of-the-art intrusion detection datasets, the CICIDS 2017 and
the UNSW-NB 2015 datasets. Experimental results showed that the proposed model signif-
icantly reduced the required training sample size and feature set size. More specifically, the
model reduced the training sample size by 74% and the feature size by up to 50%. Moreover,
hyper-parameter optimization helped improve the model performance with the detection ac-
curacies being over 99% for both datasets. This represents an improvement of 1-2% in terms
of accuracy and 1-2% in false alarm rate when compared to other works from the literature.
Salo et al. (2019) proposed an ensemble feature selection and an anomaly detection
method for network intrusion detection. The proposed framework combined unsupervised
and supervised ML techniques to classify network traffic and identify previously unseen at-
tack patterns. To that end, the authors used three different feature selection techniques that
identified 8 common and representative features. Moreover, the authors adopted k-Means
clustering to segregate the training instances and developed the classification model ac-
cordingly. Their experimental results showed that the proposed framework was effective in
detecting previously unseen attack patterns in comparison to the traditional classification
approaches.
Wang et al. (2017) proposed the use of an SVM with augmented features for their intru-
sion detection framework. More specifically, the author used the logarithm marginal density
ratios transformation to get better-quality features. Using the NSL-KDD dataset, their ex-
Machine Learning Towards Intelligent Systems: Applications, Challenges, and Opportunities 23
periments showed that the proposed framework achieved better performance in terms of
accuracy, detection rate, false alarm rate and efficiency.
5.1.3 Research Opportunities:
Although the use of ML for network intrusion detection is popular, it still requires further re-
search. One potential opportunity is to study the performance of more complex models such
as bagging ensemble models or deep learning models. This is particularly important for
real-time or near real-time network intrusion detection. Another opportunity is to study the
impact of different optimization models and techniques in enhancing the current intrusion
detection frameworks and models. A third research opportunity is to consider time-series
analysis techniques to identify and detect temporal-based anomalies and intrusion attempts.
This is crucial given that many attacks such as denial-of-service (DoS) attacks span a period
of time rather than being instantaneous. Another opportunity is to investigate the perfor-
mance of reinforcement learning and transfer learning techniques in IDSs. This is based on
the fact that such techniques have the potential to make the IDSs more flexible and effective.
Although there have been some works that have considered the use deep learning models
and time series analysis such as the work by Nguyen et al. (2020), this should be extended
to other network security problems and not just for IDSs.
5.2 Botnets Detection
5.2.1 Challenge Description:
The term botnet refers to a network of computers (bots) which have been compromised
by an attacker (aka botmaster) who has installed malicious software on the network via an
attacking technique such as trojan horses, worms and viruses. Botmasters often choose to
attack computer networks that contain many computers due to the large amounts of band-
width and powerful computing capabilities available for such networks. Once the botmaster
has control of a network, they use the network to initiate various malicious activities such
as email spam, distributed denial-of-service (DDOS) attacks, password cracking, and key
logging (Wang et al., 2015; Moubayed et al., 2020a; Injadat et al., 2020).
Zaidi and Tanveer (2017) divided the botnet life-cycle into three phases: 1. formation,
2. Command and control (C&C), 3. botnet application phase. In the formation phase, the
botmaster infects other machines on the Internet, turning them into bots on the botnet. In
the C&C phase, bots receive instructions from the bot master. During the botnet applica-
tion phase, the bots carry out malicious activities based on the instructions of the botmaster.
Although some bots might be detected and removed from the botnet, the botmaster will con-
tinue to probe the botnet in a stealthy manner for information about active bots and will plan
to form a new botnet (Vormayr et al., 2017).
One common type of botnet is the Internet relay chat (IRC) botnet. This botnet uses
IRC to facilitate command and control (C&C) communication between bots and botmas-
ters. IRC botnets can connect to one or more servers, making it easy for the botmaster to
execute commands. However, IRC botnets can be stopped by shutting-down the IRC bot-
net’s C&C server. Once attackers realized this central flaw of IRC botnets, they began to
utilize peer-to-peer (P2P) botnets. In a P2P botnet, there is no centralized server and bots
are connected to each other topologically and act as both C&C server and client. Therefore,
24 MohammadNoor Injadat et al.
even if a P2P botnet loses some of its bots, its communication will not be disrupted. Ac-
cording to Wang et al. (2015), botnets have become one of the most significant threats to the
Internet.
5.2.2 Previous Works:
McDermott et al. (2018) proposed a deep learning-based model to detect botnet activity
within IoT devices and networks. More specifically, the authors developed a Bidirectional
Long Short Term Memory based Recurrent Neural Network (BLSTM-RNN). The detec-
tion model was compared to a default LSTM-RNN. The performance was evaluated using
the accuracy and loss metrics. Experimental results demonstrated that the proposed model
achieved a detection accuracy ranging between 92%-99%, highlighting the effectiveness of
RNN-based models for botnet detection.
Pektas¸ and Acarman (2017) investigated the ability of three ML algorithms (RF, LR, and
SVM) to effectively select features to use in botnet detection during network flow analysis.
More specifically, the authors investigated three different feature selection methods. This in-
cluded Lasso linear regression models, Recursive Feature Elimination (RFE), and tree-based
feature selection, along with three different classifiers (LR, NB, and RF). The authors found
that when the meta-classifier RF was applied on the features selected by RF that the model
was nearly 99.9% accurate, making it the most accurate model that was tested. This model
almost achieved perfect classification accuracy for identifying botnet and normal traffic.
Chen et al. (2017) also discussed using ML to detect botnets. According to the authors,
in the past, signature-based and anomaly-based intrusion detection systems (IDS) were used
to detect botnets. However, as the speed of the Internet has increased, these methods are no
longer as effective. The authors proposed a method that uses conversation-based network
traffic analysis and supervised ML to identify malicious botnet traffic . The authors showed
that their approach outperformed other approaches which are based on network flow analy-
sis. More specifically, the authors’ model resulted in a 13.2% decrease in the false positive
rate of botnet traffic detection. Furthermore, it was shown that the RF algorithm had a high
detection accuracy (93.6%) and a low false positive rate (0.3%).
5.2.3 Research Opportunities:
As mentioned earlier, there are still many research opportunities in the usage of ML for bot-
net detection that are worth exploring. One such opportunity is investigating the use of hy-
brid ML models to see if they can satisfy all the requirements of their proposed online botnet
detection framework. Another potential opportunity is to consider non-numerical features as
part of any botnet detection models since such features may contain valuable information.
A third opportunity again is studying the impact of different optimization models and tech-
niques on the performance of current botnet detection models.
5.3 Intrusion Detection in Vehicles
5.3.1 Challenge Description:
In recent years, the conventional mechanical controlling parts in cars have largely been re-
placed by Electronics Control Units (ECUs) (Liu et al., 2017). ECUs are computing devices
that are used for controlling and monitoring the subsystems of a vehicle for energy efficiency
Machine Learning Towards Intelligent Systems: Applications, Challenges, and Opportunities 25
enhancement, and noise and vibration reduction (Kang and Kang, 2016; Moubayed et al.,
2020b). The use of computing devices in vehicles has led to the use of automotive network-
ing services such as Vehicle-to-Vehicle (V2V) and Vehicle-to-Infrastructure (V2I) services.
V2V automotive networking services require computing devices to perform intra-vehicular
communication, while V2I automotive networking services require computing devices to
perform inter-vehicular communication (Moubayed and Shami, 2020).
One standard communication protocol for in-vehicle network communication is Con-
troller Area Network (CAN). CAN connects sensors and actuators with ECUs (Yang et al.,
2019a). Important information such as diagnostic, informative, and controlling data is deliv-
ered through a CAN bus and it is important that this information is secured in order to keep
the driver safe. However, whenever networks are used, there is a potential for significant
security concerns. For in-vehicular networks there are several security flaws. For example,
ECUs can obtain any ECU-to-ECU broadcasting messages in the same bus, but they are
unable to identify a sender (Kang and Kang, 2016).
5.3.2 Previous Works:
Based on their concerns about the security issues of in-vehicular networks, especially the
CAN bus component, Kang and Kang (2016) created an intrusion detection system that
uses a deep neural network (DNN). The authors’ DNN was able to more accurately detect
intrusions than a traditional ANN. According to the authors, this increased accuracy is due
to the deep learning framework, which allows for the initialization of parameters through
the unsupervised pre-training of deep belief networks (DBN). Finally, using experimental
results, the authors showed that their algorithm can provide a real-time response to an attack
with a detection ratio average of 98%.
In a similar manner, Yang et al. (2019a) proposed an IDS for autonomous and connected
vehicles using DT structures. The goal of the IDS is to detect network attacks within the ve-
hicle and external to it. Experimental results showed that the authors’ proposed framework
improved the detection accuracy, detection rate, and F1 score by close to 2-3% and achieved
lower false alarm rate than other traditional methods proposed in the literature. Moreover,
the developed IDS detected various attacks. The proposed model achieved an accuracy of
100% and 99.86% on the CAN intrusion and CICIDS2017 data sets. Additionally, it reduced
the computational time by 73.7% to 325.6s and by 38.6% to 2774.8s, respectively.
Zeng et al. (2019) proposed a deep Learning-based intrusion detection model composed
of a combination of CNN and LSTM to detect malware traffic for on board units. The pro-
posed model is fed the raw traffic instead of the human-extracted private information fea-
tures. The performance of the proposed model was compared with previous methods on a
public dataset and a simulated real-life VANET dataset. Experimental results showed that
the proposed model outperformed the other methods by achieving a precision value between
95%-99% and an F1-score between 0.92-0.99.
5.3.3 Research Opportunities:
There is still ample research opportunities to integrate ML as part of IDS systems for ve-
hicular networks. For example, it is worth exploring the impact of different optimization
techniques and meta-heuristics such as particle swarm optimization and Baysian optimiza-
tion to tune the hyper-parameters of existing IDS models (Yang and Shami, 2020). This
should be done in order to improve the overall performance of such models. Another poten-
tial research opportunity is developing more complex and hybrid ML systems that can detect
26 MohammadNoor Injadat et al.
Fig. 4: Potential Deployment of ML in Network Security
Table 3: Challenges, Previous Works, and Research Opportunities within Network Security Field
Challenge Previous Work Research Opportunity
Network
Intrusion
Detection
Systems
Used a deep learning approach to detect
network intrusions (Javaid et al., 2016)
- Study the performance of more com-
plex models such as bagging ensem-
ble models, deep learning models, rein-
forcement learning, and transfer learn-
ing.
Used Bayesian optimization to hyper-
tune the parameters of three classifica-
tion algorithms for anomaly-based IDSs
(Injadat et al., 2018)
- Study the impact of different optimiza-
tion models and techniques in enhancing
the current intrusion detection frame-
works and models.
Proposed a novel multi-stage optimized
ML-based NIDS framework that re-
duced the computational complexity
while maintaining its detection perfor-
mance (Injadat et al., 2020b)
- Consider time-series analysis tech-
niques to identify and detect temporal-
based anomalies and intrusion attempts.
Proposed an ensemble feature selection
and an anomaly detection method for
network intrusion detection (Salo et al.,
2019)
- Explore the performance of NIDS
using recent datasets such as CI-
CIDS2017, CSE-CIC-IDS2018 and Ky-
oto 2006+.
Used SVM with augmented features
for their intrusion detection framework
(Wang et al., 2017)
Botnets
Detection
Proposed a BLSTM-RNN model to de-
tect botnets (McDermott et al., 2018)
- Investigate the use of hybrid ML mod-
els to see if they can satisfy all the re-
quirements of their proposed online bot-
net detection framework.
Used three ML algorithms to perform
botnet detection during network flow
analysis (Pektas¸ and Acarman, 2017).
- Study the impact of different optimiza-
tion models and techniques on current
botnet detection frameworks and mod-
els.
Used conversation-based network traf-
fic analysis and supervised ML to iden-
tify malicious botnet traffic (Chen et al.,
2017)
- Consider non-numerical features as
part of any botnet detection models
since such features may contain valu-
able information.
Intrusion
Detection in
Vehicles
Used a deep neural network (DNN) for
intrusion detection in vehicles (Kang
and Kang, 2016)
- Explore the impact of different opti-
mization techniques and meta-heuristics
to tune the hyper-parameters of existing
IDS models.
Proposed a DT-based IDS for au-
tonomous/connected vehicles (Yang
et al., 2019a)
- Develop more complex and hybrid ML
systems that can detect both the known
and unknown attacks in vehicular net-
works.
Proposed a deep Learning-based intru-
sion detection model composed of a
combination of CNN and LSTM to de-
tect malware traffic for on board units
(Zeng et al., 2019)
Machine Learning Towards Intelligent Systems: Applications, Challenges, and Opportunities 27
both the known and unknown attacks in vehicular networks. This is particularly important
since more novel attacks are being introduced that are targeting autonomous and connected
vehicles.
Table 3 summarizes the previously discussed challenges and present some of the liter-
ature work that has been conducted within this field. Moreover, they also list some of the
potential research opportunities in which ML can play a role. Also, Figure 4 provides a
visualization of how Network Intrusion Detection System (NIDS), Detecting Botnets, and
Intrusion Detection in Vehicles fit into the Detect level of the NIST’s Cybersecurity Frame-
work.
6 Banking & Finance
Moving on from network security, ML also has application in the sectors of banking and fi-
nance. After the financial crises of the 1980’s and 90’s, risk assessment of financial interme-
diaries became a hot topic. “A financial intermediary is an entity that acts as the middleman
between two parties in a financial transaction, such as a commercial bank, investment banks,
mutual funds and pension funds” (Chen, 2019). Researchers such as Chen et al. (2016) be-
lieve that ML algorithms can be used to predict individual risk in the credit portfolios of
institutions. In turn, this will help in determining who will and will not repay various forms
of credit (e.g., loans, mortgages, and credit cards). Khandani et al. echo this sentiment as
they discuss the importance of using “hard” information (e.g., characteristics contained in
consumer credit files collected by credit bureau agencies) to determine the creditworthi-
ness of consumers (Khandani et al., 2010). In the past, human discretion has been used
to determine the creditworthiness of consumers. However, ML offers a way to determine
the creditworthiness of consumers based on vast amounts of hard information. This section
will discuss how ML can be used to assess the credit risk of potential borrowers, predict if
borrowers will go bankrupt, and predict currency crises.
6.1 Credit Risk Assessment
6.1.1 Challenge Description:
With the increased dependency on mortgages and banks for financial support, credit risk
assessment has garnered significant interest from both practitioners and researchers. This is
especially crucial for financial institutions to be able to differentiate between “good” and
“bad” applicants to minimize their risk (Bao et al., 2019). This applies to both individual
applicants as well as small-medium enterprise (SMEs) applicants (Zhu et al., 2019). Multi-
ple factors are typically considered when using traditional assessment systems (Zhang et al.,
2018b). However, an applicant’s dynamic transaction history, an important indicator of the
applicant’s trustworthiness and creditworthiness, is often not considered. Additionally, in
the case of SMEs, the enterprise’s “self-oriented” factor and “supply chain finance-oriented”
factor are often neglected when assessing credit risk. Therefore, it is important for any credit
assessment system to consider multiple factors to better utilize available resources.
28 MohammadNoor Injadat et al.
6.1.2 Previous Works:
Bao et al. (2019) proposed the combined use of unsupervised and supervised ML models
to assess the credit risk of individuals. More specifically, the authors explored two different
clustering models, namely k-means and the self-organizing map (SOM) in addition to seven
potential supervised classification models including LR, DT, gradient boosting decision tree
(GBDT), RF, SVM, K-NN, and artificial neural networks (ANN). The authors studied the
performance of their proposed models using three datasets from China, Germany, and Aus-
tralia respectively. Experimental results showed that the detection accuracy of the potential
models ranged between 81%-91% for the Chinese dataset, between 64%-79% for the Ger-
man dataset, and 64%-86% for the Australian dataset.
Zhu et al. (2019) proposed a hybrid ensemble ML model to assess the credit risk of
SMEs in supply chain finance. The authors integrated two ensemble learning models, namely
the random subspace (RS) model and the multi-boosting model based on DT algorithm to
improve the performance of the credit risk assessment process. The performance of the pro-
posed model was evaluated on a Chinese dataset collected between 31 March 2014 and 31
December 2015. Experimental results showed that the assessment accuracy ranged between
67%-84% with the hybrid RS-multi-boosting model achieving the highest accuracy.
On the other hand, Xu and He (2020) proposed a deep learning model for SME credit
risk assessment. More specifically, the authors proposed a DBN composed of the Restricted
Botlzman Machine (RBM) and Softmax classifier to predict the credit risk of SMEs working
in the online supply chain space. The authors evaluated the performance of their proposed
model using three different datasets and was compared to the performance of SVM and LR
models. Experimental results showed that the proposed DBN achieved the highest accuracy
of 96% compared to 82% and 87% for the LR and SVM models respectively.
6.1.3 Research Opportunities:
Despite the literature showing that using ML has great potential for credit risk prediction,
there are still research opportunities in this field. One potential opportunity to explore is
optimizing the hyper-parameters of the ML models considered. As shown in the previous
works, most of the proposed models only consider default parameters without any attempt to
optimize them, which may result in reduced performance. Another potential opportunity is
exploring the performance of different models that can make short, medium, and long term
risk prediction rather than just on the short term. In a similar manner to the work in (Lei
et al., 2019), a third opportunity is to consider other deep learning models and architectures
such as CNN, RNN, DBNs, and Generative Adversarial Networks (GANs) to investigate the
improvement in the credit assessment accuracy.
6.2 Bankruptcy Prediction
6.2.1 Challenge Description:
When selecting potential clients one aspect of their amount of credit risk is the probabil-
ity that they will go bankrupt. Quantitative risk management systems, which are based on
ML models, can provide financial institutions with early warning signs of clients whose po-
tential business may fail (Antunes et al., 2017). Such failure can result in bankruptcy and
Machine Learning Towards Intelligent Systems: Applications, Challenges, and Opportunities 29
the client defaulting on their bank payments. In turn, this can have a devastating impact on
the firm owner, society, and the country’s overall economy (Alaka et al., 2018). This would
force governments to increase their rescue plans in order to maintain the economic growth
of the country which is a challenging task in itself. Prior work has used linear probability
and multivariate conditional probability models, the recursive partitioning algorithm, arti-
ficial intelligence, multi-criteria decision making, and mathematical programming in order
to predict a person’s amount of credit risk. However, the performance of the previously
proposed models heavily depends on the features and data collected.
6.2.2 Previous Works:
Kim et al. (2019) discuss how the financial sustainability of a company can maintain the
soundness of the state and society. They further discuss how the sustainability of financial
institutions is directly dependent on the financial sustainability of the bank’s borrowers.
Hence, it is important for financial institutions to evaluate the sustainability of their borrow-
ers, which is often done with the corporate financial distress prediction model. The authors
propose a novel hybrid SVM model that uses globally optimized SVMs (GOSVM) and
the genetic algorithm (GA) to predict potentially distressed burrowers. GOSVM optimizes
feature selection, instance selection, and kernel parameters; while GA simultaneously opti-
mizes multiple heterogeneous design factors of SVMs. The authors trained and tested their
model on real-world data from H commercial bank in Korea. The authors randomly chose
1,548 heavy industry companies, 774 of which had filed for financial distress between 1999
and 2002, and 774 which were non-bankrupt in this same time period. Experimental re-
sults showed that the proposed GOSVM model outperformed both non-SVM based models
and other SVM-based models at accurately predicting financial distress during the hold-out
phase. Based on these results, the authors concluded that their model improves the predic-
tion accuracy of conventional SVMs.
Similarly, Barboza et al. (2017) proposed the use of different ML models to predict
bankruptcy and default events of companies and institutions. More specifically, the authors
studied four models, namely SVM, DT bagging, DT boosting, and RF models in comparison
with other traditional models such as LR and ANN. Experimental results showed that the
proposed models achieved higher prediction accuracy during both the training and testing
stages. In particular, the bagging, boosting, and RF models all achieved a training accuracy
above 96% and a testing accuracy between 86%-87% as compared to the LR and ANN
methods which achieved a training accuracy between 82%-84% and a testing accuracy be-
tween 72%-76%.
Lin et al. (2019) also proposed the use of ensemble learning models in combination with
feature selection as part of their bankruptcy prediction models. To that end, the authors inves-
tigated two feature selection methods, namely information gain and genetic algorithm. The
authors also explored six ML models including LR, NB, ANN, DT, SVM, and K-NN. Ex-
perimental results illustrated that the bagging ensemble models achieved better performance
when compared to the single classifiers by having a lower false positive rate. Moreover, the
results also showed that genetic algorithm outperformed the information gain algorithm for
feature selection as it allowed the classifiers to achieve better performance.
On the other hand, Mai et al. (2019) proposed the use of deep learning models to predict
bankruptcy based on textual data in conjunction with accounting-based ratio and market-
based variables. In particular, the authors proposed a CNN-based model with word embed-
ding as part of their bankruptcy prediction model. To evaluate the performance of their pro-
posed models, the authors used a dataset consisting of 11,827 firms and 94,994 firm-years
30 MohammadNoor Injadat et al.
Fig. 5: Potential ML-based Credit Risk and Bankruptcy Assessment Framework
collected from Compustat North America, Center for Research in Security Prices (CRSP),
and the Securities Exchange Commission (SEC). Experimental results showed that the pro-
posed model outperformed other traditional models such as LR, RF, and SVM by achieving
higher prediction accuracies.
6.2.3 Research Opportunities:
Again, there are still many research opportunities that would benefit from ML to better pre-
dict bankruptcy. For example, one open area is studying the impact of other ML models and
kernels in performing the prediction. This is based on the fact that most previous work only
focused on a single kernel or a single ML model. Another research opportunity that should
be considered is investigating the performance of different optimization models and meta-
heuristics such as simulated annealing, tabu search, or particle swarm optimization to study
the potential trade-offbetween performance improvement and computational complexity. A
third research opportunity is to explore other deep learning models and architectures such as
RNN, DBN, and GANs to investigate their performance in comparison with CNN models
proposed in the literature. Figure 5 provides a potential ML-based credit risk and bankruptcy
assessment framework that can be deployed by banks and financial institutions.
6.3 Currency Crises Prediction
6.3.1 Challenge Description:
In the 90s, many countries suffered from a currency crisis wherein the value of their currency
became unstable. Europe experienced a currency crisis in 1992, Mexico in 1994, Asia in
1997-98, and Russia in 1998 (Lin et al., 2008). Therefore, the interest in developing such
systems increased in the aftermath of the 2008-09 global financial crisis (McMahon, 2019;
Basu et al., 2019). The interest stems from the fact that a currency crisis can damage the
world economy. Hence, it would be beneficial to create an early warning system in order
to prevent or at least to manage such events, particularly given the serious socio-economic
Machine Learning Towards Intelligent Systems: Applications, Challenges, and Opportunities 31
impact that such events can have. To that end, ML can play a major role as part of such
early systems given their ability to act as accurate prediction models and their promising
performance in other business-focused applications.
6.3.2 Previous Works:
McMahon (2019) proposed the use of ML models to predict crises in different currencies.
In particular, the author proposed the use of SVM model as it can overcome many of the
limitations of traditional crises prediction approaches including data non-linearity and the
variance-bias trade-off. To that end, the author applied the SVM model to data collected
between 1996 and 2014. Experimental results showed that the proposed model accuracy
predicted the majority of crises within that period from 17 emerging markets, thus illustrat-
ing its potential as a valuable tool for economists to use.
On the other hand, Xu et al. (2018) proposed the combination of RF and wavelet trans-
formation to predict currency crises. More specifically, the authors proposed the use of
discrete wavelet transformation (DWT) to systematically extract key time-based features
related to the exchange rate behavior over different time horizons. To evaluate the perfor-
mance of the proposed model the authors used a dataset containing instances about currency
crises between 1992 and 2015. Experimental results showed that the proposed RF-DWT
model achieved a high prediction accuracy between 89%-90%, outperforming the LR model
which achieved an accuracy between 84%-85%.
Similarly, Kinkyo (2020) also proposed the combination of RF and DWT for bi-annual
currency crises prediction based on the exchange market pressure (EMP) index. To that end,
the authors used a dataset covering 101 industrial and developing countries between 1994-
2018 from the International Financial Statistics of the International Monetary Fund (IMF).
Experimental results showed that the proposed model achieved a prediction accuracy of
close to 73%, outperforming other models by at least 5%. This highlighted the potential of
the proposed bi-annual forecasting model in providing guidance for both policy makers and
investors to detect currency risks.
In contrast, Alaminos et al. (2019) proposed a deep learning model to predict currency
crises. More specifically, the authors proposed the use of deep neural decision trees (DNDT)
and compared its performance to other widely adopted methodologies. To compare the per-
formance of the different models, the authors used a dataset consisting of 162 developed,
emerging, and developing countries with information between 1970–2017. Experimental
results showed the proposed model achieved a training prediction accuracy ranging between
98%-99% and a testing accuracy between 97%-99% across multiple geographical regions.
This highlights the potential of deep learning models for accurate currency crises prediction.
6.3.3 Research Opportunities:
Similar to other opportunities in the banking and finance sector, there are multiple research
opportunities in which ML can play a role for financial crises prediction. One opportunity
is investigating the performance of existing models in predicting financial crises in specific
fields rather than just at the macro/country level. For example, study the effectiveness of
existing models in predicting crises in the housing field since such models can be extremely
helpful for real-estate developers and landlords. Another potential opportunity is exploring
the effectiveness of different classification models in predicting such crises and studying
their complexity. A third potential research opportunity is exploring other deep learning
32 MohammadNoor Injadat et al.
Table 4: Challenges, Previous Works, and Research Opportunities within Banking and Finance Field
Challenge Previous Work Research Opportunity
Credit Risk
Assessment
Proposed the combined use of unsupervised and
supervised ML models to assess the credit risk of
individuals. (Bao et al., 2019)
- Explore hyper-parameter optimization
of the ML models considered to improve
their performance.
Proposed a hybrid ensemble ML model com-
posed of two ensemble learning models and
seven classification models to assess the credit
risk of SMEs in supply chain finance (Zhu et al.,
2019)
- Explore the performance of different
models that can make short, medium,
and long term risk prediction.
Proposed DBN composed of RBM and Softmax
classifier to predict the credit risk of SMEs work-
ing in the online supply chain space
- Consider other deep learning mod-
els and architectures such as CNN and
RNN to investigate the improvement in
the credit assessment accuracy.
Bankruptcy
Prediction
Compared performance of optimized SVM with
ANN and other ML techniques to predict institu-
tion bankruptcy (Kim et al., 2019)
- Study the impact of other ML mod-
els and kernels in performing the
bankruptcy prediction.
Explored four different ML models including
SVM, DT bagging, DT boosting, and RF models
for bankruptcy prediction (Barboza et al., 2017)
- Investigate the performance of dif-
ferent optimization models and meta-
heuristics to study the potential trade-off
between performance improvement and
computational complexity.
Proposed the use of ensemble learning models in
combination with feature selection as part of their
bankruptcy prediction models (Lin et al., 2019)
- Explore other deep learning models
and architectures such as RNN to inves-
tigate their performance in comparison
with CNN models proposed in the liter-
ature.
Proposed the use of CNN model to predict
bankruptcy based on textual data in conjunction
with accounting-based ratio and market-based
variables (Mai et al., 2019)
Currency
Crises
Prediction
Proposed the use of SVM model to predict cur-
rency crises (McMahon, 2019)
- Investigate the performance of exist-
ing models in predicting financial crises
in specific fields rather than just at the
macro/country level.
Proposed the combination of RF and wavelet
transformation to predict currency crises (Xu
et al., 2018)
- Explore the effectiveness of different
classification models in predicting such
crises and studying their complexity.
Proposed a combined RF-DWT model to predict
currency crises (Kinkyo, 2020)
- Explore other deep learning models
such as CNN and RNN to investigate
their performance.
Proposed the use of deep neural decision trees
(DNDT) and compared its performance to other
widely adopted methodologies (Alaminos et al.,
2019)
models such as CNN and RNN to investigate their performance given the promising results
achieved by other deep learning architectures.
Table 4 briefly summarizes the challenges, previous works, and potential research op-
portunities of ML within the banking and finance sector.
7 Social Media
Another emerging area in which ML has been playing a major role is the area of social me-
dia. Communications and Marketing Office, Tufts University (2019) defines social media
as “the means of interactions among people in which they create, share, and/or exchange
information and ideas in virtual communities and networks”. The first form of social me-
dia appeared in 1979 when USENET created a decentralized system of discussion boards
(Carvin, 2007). Since then, the Internet has advanced well beyond discussion boards where
Machine Learning Towards Intelligent Systems: Applications, Challenges, and Opportunities 33
interactions occur in text only. In the early 21st century, many websites were launched that
provide users with a platform to not only communicate via text, but also to share videos
and/or photos (Injadat et al., 2016).
According to Communications and Marketing Office, Tufts University (2019), eight of
the most popular social media platforms are Facebook, Twitter, Youtube, Vimeo, Flickr,
Instagram, Snapchat, and LinkedIn. As of 2019, there are 2.77 billion social media users
worldwide, with it being projected that in 2021 there will be 3.02 billion social media users
worldwide (Clement, 2019). There are 2.5 quintillion bytes of social media data created each
day (Marr, 2019). Furthermore, every minute of the day Snapchat users share 527,760 pho-
tos, 456,000 tweets are sent on Twitter, Instagram users post 46,740 photos, and Facebook
users post 510,000 comments and 293,000 status updates (Marr, 2019). Also, on Facebook
more than 300 million photos are uploaded per day (Marr, 2019). In conclusion, the vast
amount of data produced by social media cannot be processed by humans. Hence, social
media provides another area of opportunity for the use of ML. In this section, how ML
techniques can be applied to social media data in order to make discoveries in the fields of
pharmacovigilance, vaccine sentiment analysis, and politics will be discussed.
7.1 Pharmacovigilance
7.1.1 Challenge Description:
One way that social media data is being used is in pharmacovigilance. Pharmacovigilance
(PhV) is defined as “the science and activities relating to the detection, assessment, under-
standing, and prevention of adverse effects or any other drug-related problem” (Lezotre,
2014). Adverse effects, also known as Adverse Drug Reactions (ADRs) are harmful reac-
tions that are caused by the intake of medication (Sarker et al., 2015). ADRs have led to
millions of deaths and hospitalizations and cost nearly seventy-five billion dollars annu-
ally. Governmental agencies such as the U.S. Food and Drug Administration (FDA) and
the European Medicines Agency (EMA), along with international organizations such as
the World Health Organization (WHO) engage in pharmacovigilance by requiring manu-
facturers to report adverse events (Nikfarjam et al., 2015). These agencies also encourage
voluntary reporting by healthcare professionals and the public. However, there is no guaran-
tee that healthcare professionals or the public will report ADRs. Furthermore, when ADRs
are voluntarily reported, the information may not be timely, may be incomplete, duplicated,
under-reported or over-reported. Due to the limited quantity and lack of quality of voluntar-
ily reported ADRs, it has become necessary to supplement voluntary reports with other data
forms. For example, information about ADRs can be acquired from health-related social
networks such as DailyStrength or on social media sites such as Twitter and Facebook.
Although these sites provide a vast amount of data for potential ADR detection, it is im-
possible for a human to analyze all of the data. Hence, natural language processing (NLP)
and ML algorithms have been used to process the data (Sarker et al., 2015; Nikfarjam et al.,
2015). A survey of the literature shows that NLP techniques are commonly used to analyze
social media data for ADRs via text classification using lexicon-based approaches. Further-
more, SVM, NB, and Maximum Entropy algorithms have been used to classify text. While
these approaches provide a novel opportunity for collecting data about ADRs, there are still
many challenges to using these approaches. For example, pure lexicon-based approaches
are often impeded by consumers not using technical terms, misspelling words, using ab-
breviations, or sentence structure irregularities. Furthermore, when supervised learning ap-
34 MohammadNoor Injadat et al.
proaches are used, they require substantial amounts of data to be manually annotated, of-
ten by a domain expert. That being said, researchers have begun to use partially supervised
(semi-supervised) algorithms in order to reduce the amount of annotated data that is required
(Sloane et al., 2015).
7.1.2 Previous Works:
O’Connor et al. (2014) utilized ML on tweets in order to discover mentions of ADRs. How-
ever, in their work the authors found that false positive errors were occurring due to non-
ADR extracted terms being classified as ADRs. As an example, the authors discuss the
username TScpCancer, which was classified as an ADR even though the word cancer is be-
ing used as a name in this context.
Patki et al. (2014) used ML techniques on social media data to automatically classify
drugs into either a normal category or a blackbox category (blackbox is a category of drugs
that the FDA has identified as having serious or life-threatening safety concerns). The au-
thors’ approach showed promise at classifying social media comments as ADRs or non-
ADRs. However, their approach was only marginally successful at classifying drugs into the
normal or blackbox categories. The authors believe that they encountered this challenge due
to their limited annotated dataset. Furthermore, the authors found it challenging to distin-
guish true signals from the noisy social media text data.
Liu and Chen (2015) proposed a SVM-based framework for integrated and high-performance
patient reported adverse drug event extraction from social media. More specifically, the au-
thors used data collected from four major diabetes and heart disease forums in the United
States and applied various natural language processing models to create the lexicon-based
datasets to be fed to the SVM classifier. Experimental results showed that the proposed
model achieved a high precision ranging between 91%-94% for ADR and between 79%-
87% for medical events.
Similarly, Alimova and Tutubalina (2017) proposed the use of SVM to accurately iden-
tify ADR posted by patients on social media platforms. To that end, the authors considered
two datasets, namely the CSIRO Adverse Drug Event Corpus (CADEC) and the Twitter Cor-
pus. From these datasets, a set of context-level and entity-level features were extracted and
provided as an input to the proposed linear SVM model. Experimental results showed that
the proposed model achieved a high precision ranging between 81%-84% for the CADEC
corpus dataset and between 64%-73% for the Twitter corpus dataset.
In contrast, Cocos et al. (2017) proposed a scalable deep-learning model to analyze
and identify ADRs in social media posts. More specifically, the authors proposed the use of
RNN that labels words in an input sequence with ADR membership tags. To that end, the au-
thors used a Twitter corpus dataset to evaluate the performance of the proposed RNN-based
framework. Experimental results showed that the authors’ model outperformed other tradi-
tional models such as the baseline lexicon matching (LM) system and conditional random
field model (CRF) by achieving an F1-measure of 0.755 for ADR identification compared
to 0.63 and 0.65 for the LM and CRF models respectively.
7.1.3 Research Opportunities:
Although many previous work has utilized ML for analyzing social media posts concerning
drugs and medications, there still exist many further opportunities. One potential oppor-
tunity is examining the effectiveness of ML classification in modeling the contextual and
semantic features of tweets. Another opportunity worth exploring is enriching the ADR
Machine Learning Towards Intelligent Systems: Applications, Challenges, and Opportunities 35
lexicon datasets so that the sentiment analysis of tweets and social media posts becomes
more accurate. Another potential research opportunity is performing temporal analyses to
mine drug-ADR patterns and investigate ADRs related to the interaction of drugs taken by
patients. Also, researchers can explore more complex classification models such as hidden
markov models (HMM) to distinguish between symptoms and side-effects mentioned in
the posts. Moreover, a transfer learning model can be explored to transfer knowledge from
one classification domain to another, i.e. potentially from one drug to another or from one
platform to another.
7.2 Social Media and Vaccines
7.2.1 Challenge Description:
Another way that ML can be used to gather information from social media data is by deter-
mining people beliefs, thoughts, and feelings about various vaccines. In recent years it has
been observed that some individuals and/or groups have negative opinions about the safety
and value of vaccines, and these negative opinions are being expressed online via social
media. These negative opinions may influence some people’s decisions to receive vaccines
or to vaccinate their children (Dunn et al., 2015; Centers for Disease Control and Preven-
tion (CDC), 2019b,a; Huang et al., 2017; Du et al., 2017). In the past decade, in the United
States and other countries, there has been an increase of parents refusing to vaccinate their
children due to their concerns about the safety of vaccines. Vaccine refusal for one’s self
or one’s child can result in unnecessary harm or even death. One way that scientists and
researchers are combating the anti-vaccination movement is by analyzing social media data
with ML algorithms in order to understand how negative opinions about vaccines spread
through social media. Once these patterns are understood, scientists and researchers hope
that they can combat the spread of misinformation.
7.2.2 Previous Works:
Dunn et al. (2015) hypothesized that when Twitter users were exposed to negative opinions
about human papillomavirus (HPV) vaccines in Twitter communities that these users would
subsequently express the negative opinions that they were exposed to by re-posting simi-
lar negative opinions. In order to examine their hypothesis, the authors analyzed temporal
sequences of messages posted on Twitter (tweets) related to HPV vaccines and the social
connections between users. The researchers’ dataset was collected between October 2013
and April 2014. The dataset consisted of 83,551 tweets written in English that included
terms related to HPV vaccines. Furthermore, the social connections (N =957,865) of the
30,621 users who posted or reposted the tweets were examined to see if they also posted or
reposted such tweets. In order to analyze this large dataset, the authors utilized a supervised
ML approach to classify the tweets. This approach required the researchers to first manually
label a random sample of tweets. Then, the labeled tweets were used to train a ML classifier
to recognize similar patterns in the remaining tweets. More specifically, the classifier was
an ensemble of four classifiers that used the content of the tweets (the words and word com-
binations in the tweets themselves) or the social relations between users (the users followed
by the user responsible for the tweet) in order to classify the sentiment of the tweets. The
sentiment of the users’ tweets about HPV vaccines was classified either as negative or neu-
tral/positive. When the four classifiers were trained and tested in a 10-fold cross validation,
36 MohammadNoor Injadat et al.
their accuracy ranged between 87.6% and 94.0%. The researchers concluded that Twitter
users who were more often exposed to negative opinions about HPV vaccines were more
likely to subsequently post negative tweets about HPV vaccines.
Similarly, Du et al. (2017) proposed a hierarchical SVM-based model to predict tweet
sentiments about HPV vaccines. To that end, the authors collected tweets written in En-
glish containing HPV vaccines-related keywords in the time between November 2, 2015
and March 28, 2016. Experimental results showed that the proposed model achieved a high
precision ranging between 71%-78%. Moreover, the model also showed particularly high
precision in identifying negative sentiments pertaining to HPV vaccine safety with a value
of around 80%.
Huang et al. (2017) used natural language classifiers to examine and analyze data from
Twitter in order to track flu vaccinations over time, as well as by geography and gender. The
researchers collected a dataset of 1,007,582 tweets. From this dataset, the researchers cre-
ated a training dataset by annotating a random sample of 10,000 tweets. After testing various
classifiers, the researchers chose the best-performing classifier, namely LR, and used it in
the rest of their experiments. When the researchers compared the results of their algorithm
to a published government survey data about vaccination from the US Centers for Disease
Control and Prevention (CDC), they found that their results were highly correlated with the
CDC’s data (r =0.90). These results suggest that ML algorithms can be applied to Twitter
data in order to track people’s attitudes and behaviors about flu vaccinations.
7.2.3 Research Opportunities:
As evident by the different research works discussed above, ML has great potential as it can
be used to examine large amounts of social media data in order to track and determine how
social media users may influence each other’s opinions of vaccines. While these studies have
shown great potential for the use of ML, there are still some limitations that offer possibilities
for future research. One limitation that could use further development is that social media
users’ connections change over time and this may not be reflected in data that is taken
from a set time period. Therefore, it is important to develop adaptive ML models that can
change with the social connection changes of the social media platform. Another potential
opportunity is creating new datasets with updated vocabulary to better track the sentiment of
users based on the non-standard abbreviations, slang and phrases commonly used on social
media platforms. A third research opportunity worth exploring is to study the performance
of deep learning architectures such as CNNs and RNNs for sentiment analysis of various
vaccines. This is particularly important given the promising performance illustrated by such
architectures in analyzing social media posts.
7.3 Social Media and Politics
7.3.1 Challenge Description:
Moving beyond health-related topics, ML techniques can also be applied to social media in
order to collect, monitor, analyze, summarize, and visualize politically relevant information
(Hosni and Li, 2019). In recent years, social media platforms such as Twitter and Facebook
have been used to increase political participation. For example, social media users publicly
spread information about their political opinions on Twitter and political institutions have
begun to use Facebook pages or groups to engage with citizens (Stieglitz and Dang-Xuan,
Machine Learning Towards Intelligent Systems: Applications, Challenges, and Opportunities 37
2013). Furthermore, politicians and political parties are interested in social media data, be-
cause they can benefit from understanding what the public thinks about them (Maynard
et al., 2012). Due to their interest in public opinion about politics, politicians or political
parties may monitor social media data in order to detect social media content that is directly
or indirectly associated with them. Furthermore, the monitoring of political social media
data is also important because it may provide information about potential political crises or
scandals. Additionally, the spread of political information through social networks can lead
to administrative, political and societal changes. For example, social media played a central
role in shaping political debates in the Arab Spring (a series of pro-democracy protests, up-
risings, and armed rebellions that spread across North Africa and the Middle East beginning
in the spring of 2011 (Editors of History.com Website, 2018)).
A common method that is used for the detection and analysis of political social media
content is opinion mining (also known as sentiment analysis). The process of political opin-
ion mining consists of collecting text that contains political opinions (or sentiments) and
extracting attributes and components about a specific political feature from said text, then
determine whether the text is positive, negative or neutral.
7.3.2 Previous Works:
Jahanbakhsh and Moon (2014) have implemented sentiment analysis on tweets. They were
interested in the predictive power of social media. In their study, the authors analyzed 32
million tweets related to the 2012 US presidential election using a combination of ML tech-
niques. The authors implemented a Twitter crawler from September 29, 2012 until Novem-
ber 16, 2012 using keywords such as Barack Obama, Mitt Romney, US election, Paul Ryan,
and Joe Biden. Their results were numerous. Firstly, the authors’ results (that Obama was
leading in Twitter for the 2012 US presidential election) matched with the outcome of the
election. Secondly, the authors found that by analyzing geo-tweets (tweets with a geo-tag)
with geographical sentiment analysis, they were able to uncover the popularity of candi-
dates across the US states. Thirdly, the authors work demonstrated that LDA is a powerful
unsupervised algorithm when combined with the NB classifier as it was able to “predict” the
result of the 2012 US election. Hence, the authors have presented a system of mining social
media data that may be used for predicting future events.
In a similar fashion, Ramteke et al. (2016) proposed the use of ML models to predict the
results of the 2016 US presidential election based on sentiment analysis of the correspond-
ing tweets. To that end, the authors proposed the use of NB and SVM models to classify the
tweets about Donald Trump and Hilary Clinton. The performance of the model was evalu-
ated using a dataset collected through Twitter between March 16th-17th, 2016 which was
later labeled manually. Experimental results showed that the proposed models achieved a
sentiment prediction accuracy ranging between 97%-99% with an F1-score between 0.94-
0.97, highlighting the potential of ML models in predicting election results based on the
sentiment analysis of tweets.
Oyebode and Orji (2019) also proposed the use of ML models to predict the results of the
2019 Nigerian presidential election by comparing three lexicon-based classifiers (VADER,
VADER-EXT, and Textblob) and five ML-based classifiers (SVM, LR, NB, stochastic gra-
dient descent SGD, and RF). To evaluate the performance of the different models, the au-
thors collected 118,421 posts between January 1 and February 22, 2019. Experimental re-
sults showed that the VADER-Ext approach outperformed the other two lexicon-based ap-
proaches by achieving a precision of approximately 81%. In a similar fashion, it was shown
that the LR method achieved the highest accuracy and precision of 77% and 78% respec-
38 MohammadNoor Injadat et al.
Fig. 6: Potential ML-based Social Media Analytics Framework
tively among the different ML models.
On the other hand, Tsai et al. (2019) extended the concept of sentiment analysis by using
deep learning models to predict multiple local election results rather than the national elec-
tion. To that end, the authors proposed the use of recursive neural tensor network (RNTN)
to analyze the sentiment shown in various Twitter posts about the 2018 US Midterm elec-
tions. The performance of the proposed model was evaluated using a manually collected
dataset consisting of approximately 800 tweets. Experimental results showed that the pro-
posed model achieved a high prediction accuracy as it predicted an advantage of 9.2% for the
Democratic candidate compared to the actual advantage which was measured to be 8.6%.
As such, it was shown that the proposed model indeed has great potential in accurately
predicting multiple local election results.
7.3.3 Research Opportunities:
Again, there are still many research opportunities in which ML can play a role as part of
a politics sentiment analysis frameworks. One such opportunity is investigating other ML
algorithms such as SVM and ANN given their previous success in determining linguistic
features for opinion classification. Another potential opportunity is collecting more features
such as swear words, sarcasm, and negative and conditional detection as well as contextual
clues features to make the sentiment analysis framework more accurate and effective. A third
opportunity is to consider other deep learning models such as CNN and RNN to compare
their performance with the currently proposed deep learning models. This is crucial given
the promising results achieved by deep learning models such as RNTN model.
Table 5 summarizes some of the different challenges and research opportunities of ML
within the social media field. Moreover, Figure 6 provides a visualization of how these topics
fit into a Social Media Analytics Framework.
Machine Learning Towards Intelligent Systems: Applications, Challenges, and Opportunities 39
Table 5: Challenges, Previous Works, and Research Opportunities within Social Media Field
Challenge Previous Work Research Opportunity
Pharmacovigilance
Utilized ML on tweets in order to dis-
cover mentions of ADRs (O’Connor
et al., 2014)
- Examine the effectiveness of ML clas-
sification in modeling the contextual and
semantic features of tweets.
Used ML techniques on social media
data to automatically classify drugs into
either a normal category or a blackbox
category (Patki et al., 2014)
- Enrich the ADR lexicon datasets so
that the sentiment analysis of tweets and
social media posts become more accu-
rate.
Proposed a SVM-based framework for
integrated and high-performance patient
reported adverse drug event extraction
from social media (Liu and Chen, 2015)
- Perform temporal analyses to mine
drug-ADR patterns and investigate
ADRs related to the interaction of drugs
taken by patients.
Proposed the use of SVM to accurately
identify ADR posted by patients on so-
cial media platforms (Alimova and Tu-
tubalina, 2017)
- Explore more complex classification
models to distinguish between symp-
toms and side-effects mentioned in the
posts.
Proposed a scalable RNN model to ana-
lyze and identify ADRs in social media
posts (Cocos et al., 2017)
- Explore transfer learning models to
transfer knowledge from one classifica-
tion domain to another.
Social Media and
Vaccines
Utilized a supervised ML approach to
classify vaccine-related tweets (Dunn
et al., 2015)
- Develop adaptive ML models that
can change with the social connection
changes of the social media platform.
Used SVM model on Twitter data in
order to assess HPV vaccination senti-
ments (Du et al., 2017)
- Create new datasets with updated vo-
cabulary to better track the sentiment of
users based on the non-standard abbre-
viations, slang and phrases commonly
used on social media platforms.
Used natural language classifiers to ex-
amine and analyze data from Twitter in
order to track flu vaccinations over time,
geography, and gender (Huang et al.,
2017)
- Study and compare the performance
of deep learning architectures such as
CNNs and RNNs.
Social Media and
Politics
Created a ML-NLP engine that imple-
mented a NB classifier for sentiment
analysis, and the Latent Dirichlet Allo-
cation (LDA) algorithm for topic mod-
eling (Jahanbakhsh and Moon, 2014)
- Investigate other ML algorithms such
as SVM and ANN given their previous
success in determining linguistic fea-
tures for opinion classification.
Proposed the use of NB and SVM mod-
els to predict the results of the 2016
US presidential election based on corre-
sponding tweets (Ramteke et al., 2016)
- Collect more features such as swear
words, sarcasm, and negative and con-
ditional detection as well as contextual
clues features to make the sentiment
analysis framework more accurate and
effective.
Compared the performance of three
lexicon-based and five ML-based mod-
els in predicting the results of the 2019
Nigerian presidential election (Oyebode
and Orji, 2019)
- Consider other deep learning mod-
els such as CNN and RNN to com-
pare their performance with the cur-
rently proposed deep learning models.
Proposed the use of RNTN deep learn-
ing model to predict the results of the
2018 US Midterm elections (Tsai et al.,
2019)
8 Conclusion
The availability and popularity of the Internet and related technologies has resulted in large
amounts of data being available for analyses. However, humans do not possess the cognitive
capabilities to understand such large amounts of data. Machine learning (ML) provides a
way for humans to process large amounts of data and come to conclusions about the data.
ML has applications in various fields. This review focused on some of the fields and ap-
40 MohammadNoor Injadat et al.
Fig. 7: Summary of Challenges and ML Techniques
plications such as education, healthcare, network security, banking and finance, and social
media. These fields each have their own unique challenges. However, ML can provide solu-
tions to these challenges, as well as create further research opportunities. Accordingly, this
work briefly described some of the challenges facing the aforementioned fields and surveyed
some of the previous literature works that focused on them. Moreover, it presented several
research opportunities on the role and potential of using ML to address these challenges. Fig-
ure 7 summarizes the challenges and previous/potential ML techniques that addressed/can
address them respectively.
Acknowledgments This study was funded by Ontario Graduate Scholarship (OGS) Program.
Conflict of Interest The authors declare that they have no conflict of interest.
Informed Consent This study does not involve any experiments on animals.
References
A Flyvbjerg CC G Holt, Goldstein B (2010) Textbook of Diabetes: A Clinical Approach, 4th edn. Wiley, New
Jersey, USA
Agarap AFM (2018) On breast cancer detection: an application of machine learning algorithms on the wis-
consin diagnostic dataset. In: Proceedings of the 2nd International Conference on Machine Learning and
Soft Computing, pp 5–9
Aher SB, Lobo L (2013) Combination of machine learning algorithms for recommendation of courses in
e-learning system based on historical data. Knowledge-Based Systems 51:1–14
Alaka HA, Oyedele LO, Owolabi HA, Kumar V, Ajayi SO, Akinade OO, Bilal M (2018) Systematic review of
bankruptcy prediction models: Towards a framework for tool selection. Expert Systems with Applications
94:164–184
Alaminos D, Becerra-Vicario R, Fern´
andez-G´
amez M ´
A, Cisneros Ruiz AJ (2019) Currency crises prediction
using deep neural decision trees. Applied Sciences 9(23):5227
Machine Learning Towards Intelligent Systems: Applications, Challenges, and Opportunities 41
Albisser AM (2003) Analysis: Toward algorithms in diabetes self-management. Diabetes technology & ther-
apeutics 5(3):371–373
Alimova I, Tutubalina E (2017) Automated detection of adverse drug reactions from social media posts
with machine learning. In: International Conference on Analysis of Images, Social Networks and Texts,
Springer, pp 3–15
Antunes F, Ribeiro B, Pereira F (2017) Probabilistic modeling and visualization for bankruptcy prediction.
Applied Soft Computing 60:831–843
Bao W, Lianju N, Yue K (2019) Integration of unsupervised and supervised machine learning algorithms for
credit risk assessment. Expert Systems with Applications 128:301–315
Barboza F, Kimura H, Altman E (2017) Machine learning models and bankruptcy prediction. Expert Systems
with Applications 83:405–417
Barron-Estrada ML, Zatarain-Cabada R, Oramas-Bustillos R, Gonzalez-Hernandez F (2017) Sentiment anal-
ysis in an affective intelligent tutoring system. In: 2017 IEEE 17th International Conference on Advanced
Learning Technologies (ICALT), pp 394–397
Basu SS, Perrelli RA, Xin W (2019) External crisis prediction using machine learning: Evidence from three
decades of crises around the world. Computing in Economics and Finance, Ottawa, Canada
Bawa P (2016) Retention in online courses: Exploring issues and solutions a literature review. Sage Open
6(1)
Bellazzi R (2008) Telemedicine and diabetes management: current challenges and future research directions.
Journal of diabetes science and technology 2(1):98–104
Bhatia K, Arora S, Tomar R (2016) Diagnosis of diabetic retinopathy using machine learning classification
algorithm. In: 2016 2nd International Conference on Next Generation Computing Technologies (NGCT),
pp 347–351
Black G (2002) A comparison of traditional, online, and hybrid methods of course delivery. Journal of Busi-
ness Administration Online 1(1):1–9
Bonawitz K, Eichner H, Grieskamp W, Huba D, Ingerman A, Ivanov V, Kiddon C, Konecn `
y J, Mazzocchi S,
McMahan HB, et al. (2019) Towards federated learning at scale: System design. In: 2nd Conference on
Machine Learning and Systems (SysML 2019)
Bourkoukou O, El Bachari E (2016) E-learning personalization based on collaborative filtering and learner’s
preference. Journal of Engineering Science and Technology 11(11):1565–1581
Bourkoukou O, El Bachari E (2018) Toward a hybrid recommender system for e-learning personnalization
based on data mining techniques. JOIV: International Journal on Informatics Visualization 2(4):271–278
Bughin J, Seong J, Manyika J, H¨
am¨
al¨
ainen L, Windhagen E, Hazan E (2019) Notes from the ai frontier:
Tackling europe?s gap in digital and ai. McKinsey&Company: New York, NY, USA
Caban JJ, Gotz D (2015) Visual analytics in healthcare–opportunities and research challenges
Caldas S, Meher Karthik Duddu S, Wu P, Li T, Kone ˇ
cn`
y J, McMahan HB, Smith V, Talwalkar A (2019)
Leaf: A benchmark for federated settings. In: Workshop on Federated Learning for Data Privacy and
Confidentiality
Carvin A (2007) Timeline: The life of the blog. URL https://www.npr.org/templates/story/story.
php?storyId=17421022, Accessed on: Jan. 5, 2020
Centers for Disease Control and Prevention (CDC) (2019a) Attention adults: You need vaccines too! URL
https://www.cdc.gov/features/adultimmunizations/index.html, Accessed on: Jan. 13, 2020
Centers for Disease Control and Prevention (CDC) (2019b) If you choose not to vaccinate your
child, understand the risk and responsibilities. URL https://www.cdc.gov/vaccines/parents/
vaccine-decision/no- vaccination.html, Accessed on: Jan. 13, 2020
Chanoch LH, Jovanovic L, Peterson CM (1985) The evaluation of a pocket computer as an aid to insulin dose
determination by patients. Diabetes Care 8(2):172–176
Chen J (2019) Financial intermediary. URL https://www.investopedia.com/terms/f/
financialintermediary.asp, Accessed on: Nov. 1, 2019
Chen N, Ribeiro B, Chen A (2016) Financial credit risk assessment: a recent review. Artificial Intelligence
Review 45(1):1–23
Chen R, Niu W, Zhang X, Zhuo Z, Lv F (2017) An effective conversation-based botnet detection method.
Mathematical Problems in Engineering 2017
Chiarelli F, Tumini S, Morgese G, Albisser AM (1990) Controlled study in diabetic children comparing
insulin-dosage adjustment by manual and computer algorithms. Diabetes Care 13(10):1080–1084
Chiu YC, Chen HIH, Zhang T, Zhang S, Gorthi A, Wang LJ, Huang Y, Chen Y (2019) Predicting drug
response of tumors from integrated genomic profiles by deep neural networks. BMC medical genomics
12(1):18
Chung JY, Lee S (2019) Dropout early warning systems for high school students using machine learning.
Children and Youth Services Review 96:346–353
42 MohammadNoor Injadat et al.
Cisco (2019) What is network security? URL https://www.cisco.com/c/en/us/products/security/
what-is- network-security.html, Accessed on: Feb. 1, 2020
Clement J (2019) Number of social network users worldwide from 2010 to 2021. URL https://
www.statista.com/statistics/278414/number-of- worldwide-social-network- users/, Ac-
cessed on: Dec. 1, 2019
Cocos A, Fiks AG, Masino AJ (2017) Deep learning for pharmacovigilance: recurrent neural network archi-
tectures for labeling adverse drug reactions in twitter posts. Journal of the American Medical Informatics
Association 24(4):813–821
Columbus L (2020) Roundup of machine learning forecasts and market estimates, 2020. Forbes
Communications and Marketing Office, Tufts University (2019) Social media overview. URL
https://communications.tufts.edu/marketing-and- branding/social-media- overview/,
Accessed on: Jan. 19, 2020
Coussement K, Phan M, De Caigny A, Benoit DF, Raes A (2020) Predicting student dropout in subscription-
based online learning environments: The beneficial impact of the logit leaf model. Decision Support Sys-
tems p 113325
Di Pietro R, Distefano S (2019) An intelligent tutoring system tool combining machine learning and gamifi-
cation in education. In: TOOLS: International Conference on Objects, Components, Models and Patterns,
Springer International Publishing, pp 218–226
Dong D, Zhang W, Jing Q (2019) Paddle Federated Learning. Available at:
https://paddlefl.readthedocs.io/en/latest/introduction.html
Du J, Xu J, Song HY, Tao C (2017) Leveraging machine learning-based approaches to assess human papil-
lomavirus vaccination sentiment trends with twitter data. BMC medical informatics and decision making
17(2):69
Dunn AG, Leask J, Zhou X, Mandl KD, Coiera E (2015) Associations between exposure to and expression of
negative opinions about human papillomavirus vaccines on social media: an observational study. Journal
of medical Internet research 17(6):e144
Dwivedi P, Bharadwaj KK (2015) e-learning recommender system for a group of learners based on the unified
learner profile approach. Expert Systems 32(2):264–276
Editors of Historycom Website (2018) Arab spring. URL https://www.history.com/topics/
middle-east/arab- spring, Accessed on: Jan. 21, 2020
Elfaki AO, Alhawiti KM, AlMurtadha YM, Abdalla OA, Elshiekh AA (2014) Rule-based recommendation
for supporting student learning-pathway selection. Recent Advances in Electrical Engineering and Educa-
tional Technologies pp 155–160
Felder RM, Silverman LK, et al. (1988) Learning and teaching styles in engineering education. Engineering
education 78(7):674–681
Forum WE (2018) The future of jobs report 2018. World Economic Forum Geneva
Gadekallu TR, Khare N, Bhattacharya S, Singh S, Reddy Maddikunta PK, Ra IH, Alazab M (2020) Early
detection of diabetic retinopathy using pca-firefly based deep learning model. Electronics 9(2):274
Halford GS, Baker R, McCredden JE, Bain JD (2005) How many variables can humans process? Psycholog-
ical science 16(1):70–76
Harvard Medical School (2017) Retinopathy. Available at: https://www.health.harvard.edu/a_to_z/
retinopathy-a- to-z
Herrero P, Pesl P, Reddy M, Oliver N, Georgiou P, Toumazou C (2014) Advanced insulin bolus advisor
based on run-to-run control and case-based reasoning. IEEE journal of biomedical and health informatics
19(3):1087–1096
Holzinger A, Dehmer M, Jurisica I (2014) Knowledge discovery and interactive data mining in
bioinformatics-state-of-the-art, future challenges and research directions. BMC bioinformatics 15(6):I1
Hosni AIE, Li K (2019) Minimizing the influence of rumors during breaking news events in online social
networks. Knowledge-Based Systems p 105452
Huang C, Clayton EA, Matyunina LV, McDonald LD, Benigno BB, Vannberg F, McDonald JF (2018) Ma-
chine learning predicts individual cancer patient responses to therapeutic drugs with high accuracy. Scien-
tific reports 8(1):1–8
Huang X, Smith MC, Paul MJ, Ryzhkov D, Quinn SC, Broniatowski DA, Dredze M (2017) Examining
patterns of influenza vaccination in social media. In: Workshops at the Thirty-First AAAI Conference on
Artificial Intelligence
Hussain L, Ahmed A, Saeed S, Rathore S, Awan IA, Shah SA, Majid A, Idris A, Awan AA (2018) Prostate
cancer detection using machine learning techniques by employing combination of features extracting
strategies. Cancer Biomarkers 21(2):393–413
Ingerman A, Ostrowski K (2019) Introducing tensorflow federated. Available at: https://medium.
com/tensorflow/introducing-tensorflow-federated-a4147aa20041
Machine Learning Towards Intelligent Systems: Applications, Challenges, and Opportunities 43
Injadat M, Salo F, Nassif AB (2016) Data mining techniques in social media: A survey. Neuro-
computing 214:654 – 670, DOI https://doi.org/10.1016/j.neucom.2016.06.045, URL http://www.
sciencedirect.com/science/article/pii/S092523121630683X
Injadat M, Salo F, Nassif AB, Essex A, Shami A (2018) Bayesian optimization with machine learning algo-
rithms towards anomaly detection. In: 2018 IEEE Global Communications Conference (GLOBECOM),
pp 1–6, DOI 10.1109/GLOCOM.2018.8647714
Injadat M, Moubayed A, Nassif AB, Shami A (2020a) Multi-split Optimized Bagging Ensemble
Model Selection for Multi-class Educational Datasets. Applied Intelligence DOI https://doi.org/10.1007/
s10489-020- 01776-3
Injadat M, Moubayed A, Nassif AB, Shami A (2020b) Multi-Stage Optimized Machine Learning Framework
for Network Intrusion Detection. IEEE Transactions On Network and Service Management
Injadat M, Moubayed A, Nassif AB, Shami A (2020c) Systematic ensemble model selec-
tion approach for educational data mining. Knowledge-Based Systems 200:105992, DOI https://
doi.org/10.1016/j.knosys.2020.105992, URL http://www.sciencedirect.com/science/article/
pii/S0950705120302999
Injadat M, Moubayed A, Shami A (2020) Detecting Botnet Attacks in IoT Environments: An Optimized
Machine Learning Approach. In: IEEE 32nd International Conference on Microelectronics (ICM2020)
Jahanbakhsh K, Moon Y (2014) The predictive power of social media: On the predictability of us presidential
elections using twitter. arXiv preprint arXiv:14070622
Javaid A, Niyaz Q, Sun W, Alam M (2016) A deep learning approach for network intrusion detection system.
In: Proceedings of the 9th EAI International Conference on Bio-inspired Information and Communications
Technologies (formerly BIONETICS), pp 21–26
Jelinek HF, Stranieri A, Yatsko A, Venkatraman S (2016) Data analytics identify glycated haemoglobin co-
markers for type 2 diabetes mellitus diagnosis. Computers in biology and medicine 75:90–97
Jovanovic L, Peterson CM (1982) Optimal insulin delivery for the pregnant diabetic patient. Diabetes Care
5:24–37
Kang MJ, Kang JW (2016) Intrusion detection system using deep neural network for in-vehicle network
security. PloS one 11(6)
Kavakiotis I, Tsave O, Salifoglou A, Maglaveras N, Vlahavas I, Chouvarda I (2017) Machine learning and
data mining methods in diabetes research. Computational and structural biotechnology journal 15:104–
116
Kearns MJ, Vazirani UV, Vazirani U (1994) An introduction to computational learning theory. MIT press
Khandani AE, Kim AJ, Lo AW (2010) Consumer credit-risk models via machine-learning algorithms. Journal
of Banking & Finance 34(11):2767–2787
Kim Kj, Lee K, Ahn H (2019) Predicting corporate financial sustainability using novel business analytics.
Sustainability 11(1):64
Kinkyo T (2020) A bi-annual forecasting model of currency crises. Applied Economics Letters 27(4):255–
261
Klaˇ
snja-Mili´
cevi´
c A, Vesin B, Ivanovi´
c M, Budimac Z (2011) E-learning personalization based on hybrid
recommendation strategy and learning style identification. Computers & Education 56(3):885–899
Konecn`
y J, McMahan HB, Felix XY, Richt´
arik P, Suresh AT, Bacon D (2016) Federated learning: Strategies
for improving communication efficiency. In: 29th Conference on Neural Information Processing Systems
(NIPS 2016)
Lei K, Xie Y, Zhong S, Dai J, Yang M, Shen Y (2019) Generative adversarial fusion network for class
imbalance credit scoring. Neural Computing and Applications pp 1–12
Lezotre PL (2014) Part iii - recommendations to support the next phase of international cooperation, conver-
gence, and harmonization in the pharmaceutical domain. In: Lezotre PL (ed) International Cooperation,
Convergence and Harmonization of Pharmaceutical Regulations, Academic Press, Boston, pp 221 – 294,
DOI https://doi.org/10.1016/B978-0-12-800053- 3.00004-5, URL http://www.sciencedirect.com/
science/article/pii/B9780128000533000045
Lin CS, Khan HA, Chang RY, Wang YC (2008) A new approach to modeling early warning systems for cur-
rency crises: Can a machine-learning fuzzy expert system predict the currency crises effectively? Journal
of International Money and Finance 27(7):1098–1121
Lin WC, Lu YH, Tsai CF (2019) Feature selection in single and ensemble learning-based bankruptcy predic-
tion models. Expert Systems 36(1):e12335
Liu X, Chen H (2015) A research framework for pharmacovigilance in health social media: identification and
evaluation of patient adverse drug event reports. Journal of biomedical informatics 58:268–279
Liu X, Zhang P, Wang F, Hu Y, Liu H (2017) Research on automotive brake-by-wire system based on flexray
bus. In: 2017 5th International Conference on Frontiers of Manufacturing Science and Measuring Tech-
nology (FMSMT 2017), Atlantis Press
44 MohammadNoor Injadat et al.
Mahana M, Johns M, Apte A (2012) Automated essay grading using machine learning. Mach Learn Session,
Stanford University
Mai F, Tian S, Lee C, Ma L (2019) Deep learning models for bankruptcy prediction using textual disclosures.
European journal of operational research 274(2):743–758
Marr B (2019) How much data do we create every day? the mind-blowing stats every-
one should read. URL https://www.forbes.com/sites/bernardmarr/2018/05/21/
how-much- data-do- we-create-every- day-the- mind-blowing- stats-everyone-should- read/
#259844e160ba, Accessed on: Nov. 20, 2019
Mathias S, Bhattacharyya P (2018) ASAP++: Enriching the ASAP automated essay grading dataset with
essay attribute scores. In: Proceedings of the Eleventh International Conference on Language Resources
and Evaluation (LREC 2018)
Mathias S, Bhattacharyya P (2020) Can Neural Networks Automatically Score Essay Traits? In: Proceedings
of the Fifteenth Workshop on Innovative Use of NLP for Building Educational Applications, Association
for Computational Linguistics, pp 85–91, DOI 10.18653/v1/2020.bea-1.8
Maynard D, Bontcheva K, Rout D (2012) Challenges in developing opinion mining tools for social media.
In: Proceedings of the Language Resources and Evaluation Conference (LREC), pp 15–22
McDermott CD, Majdani F, Petrovski AV (2018) Botnet detection in the internet of things using deep learning
approaches. In: 2018 International Joint Conference on Neural Networks (IJCNN), pp 1–8
McMahon MJ (2019) Rethinking early warning systems: Using the radial based support vector machine to
forecast currency crises. PhD thesis, Claremont Graduate University
Michael Dansinger (2019) What is a glycated hemoglobin test (hba1c)? URL https://www.webmd.com/
diabetes/qa/what-is- a-glycated- hemoglobin-test-hba1c
Miotto R, Wang F, Wang S, Jiang X, Dudley JT (2018) Deep learning for healthcare: review, opportunities
and challenges. Briefings in bioinformatics 19(6):1236–1246
Mitchell TM (1997) Machine Learning. McGraw-Hill, New York, NY, USA
Mondal B, Patra O, Mishra S, Patra P (2020) A course recommendation system based on grades. In: 2020
International Conference on Computer Science, Engineering and Applications (ICCSEA), pp 1–5
Moocorg (2019) Massive open online courses: An edx site. URL https://www.mooc.org/, Accessed on:
Dec. 15, 2019
Moubayed A (2018) Optimization Modeling and Machine Learning Techniques Towards Smarter Systems
and Processes. PhD thesis, University of Western Ontario
Moubayed A, Shami A (2020) Softwarization, virtualization, & machine learning for intelligent & effective
v2x communications. IEEE Intelligent Transportation Systems Magazine
Moubayed A, Injadat M, Nassif AB, Lutfiyya H, Shami A (2018) E-learning: Challenges and research op-
portunities using machine learning data analytics. IEEE Access 6:39117–39138, DOI 10.1109/ACCESS.
2018.2851790
Moubayed A, Injadat M, Shami A, Lutfiyya H (2018) Dns typo-squatting domain detection: A data analytics
& machine learning based approach. In: 2018 IEEE Global Communications Conference (GLOBECOM),
IEEE, pp 1–7
Moubayed A, Injadat M, Shami A, Lutfiyya H (2018) Relationship between student engagement and per-
formance in e learning environment using association rules. In: 2018 IEEE World Engineering Education
Conference (EDUNINE), pp 1–6, DOI 10.1109/EDUNINE.2018.8451005
Moubayed A, Injadat M, Shami A, Lutfiyya H (2019) Student engagement level in e learning environment:
Clustering using k means. American Journal of Distance Education 34(2), DOI 10.1080/08923647.2020.
1696140
Moubayed A, Aqeeli E, Shami A (2020) Ensemble-based Feature Selection and Classification Model for
DNS Typo-squatting Detection. In: 33rd Canadian Conference on Electrical and Computer Engineering
(CCECE’20), IEEE, pp 1–6
Moubayed A, Injadat M, Shami A (2020a) Optimized Random Forest Model for Botnet Detection Based on
DNS Queries. In: IEEE 32nd International Conference on Microelectronics (ICM2020)
Moubayed A, Shami A, Heidari P, Larabi A, Brunner R (2020b) Edge-enabled v2x service placement for
intelligent transportation systems. IEEE Transactions on Mobile Computing
Mucaki EJ, Zhao JZ, Lizotte DJ, Rogan PK (2019) Predicting responses to platin chemotherapy agents with
biochemically-inspired machine learning. Signal transduction and targeted therapy 4(1):1–12
M¨
uller PL, Treis T, Odainic A, Pfau M, Herrmann P, Tufail A, Holz FG (2020) Prediction of function in
abca4-related retinopathy using ensemble machine learning. Journal of Clinical Medicine 9(8):2428
Nguyen G, Dlugolinsky S, Tran V, Lopez Garcia A (2020) Deep learning for proactive network monitoring
and security protection. IEEE Access 8:19696–19716, DOI 10.1109/ACCESS.2020.2968718
Nikfarjam A, Sarker A, O’connor K, Ginn R, Gonzalez G (2015) Pharmacovigilance from social media:
mining adverse drug reaction mentions using sequence labeling with word embedding cluster features.
Machine Learning Towards Intelligent Systems: Applications, Challenges, and Opportunities 45
Journal of the American Medical Informatics Association 22(3):671–681
O’Connor K, Pimpalkhute P, Nikfarjam A, Ginn R, Smith KL, Gonzalez G (2014) Pharmacovigilance on
twitter? mining tweets for adverse drug reactions. In: AMIA annual symposium proceedings, American
Medical Informatics Association, vol 2014, pp 924–933
Owens C, Zisser H, Jovanovic L, Srinivasan B, Bonvin D, Doyle FJ (2006) Run-to-run control of blood
glucose concentrations for people with type 1 diabetes mellitus. IEEE Transactions on Biomedical Engi-
neering 53(6):996–1005
Oyebode O, Orji R (2019) Social media and sentiment analysis: The nigeria presidential election 2019. In:
2019 IEEE 10th Annual Information Technology, Electronics and Mobile Communication Conference
(IEMCON), pp 0140–0146
Patki A, Sarker A, Pimpalkhute P, Nikfarjam A, Ginn R, O’Connor K, Smith K, Gonzalez G (2014) Mining
adverse drug reaction signals from social media: going beyond extraction. Proceedings of BioLinkSig
2014:1–8
Pektas¸ A, Acarman T (2017) Effective feature selection for botnet detection based on network flow analysis
Prasad V, Fojo T, Brada M (2016) Precision oncology: origins, optimism, and potential. The Lancet Oncology
17(2):e81–e86
Ramalingam V, Pandian A, Chetry P, Nigam H (2018) Automated essay grading using machine learning
algorithm. In: Journal of Physics: Conference Series, IOP Publishing, vol 1000, p 012030
Ramteke J, Shah S, Godhia D, Shaikh A (2016) Election result prediction using twitter sentiment analysis.
In: 2016 International Conference on Inventive Computation Technologies (ICICT), vol 1, pp 1–5
Reddy GT, Bhattacharya S, Siva Ramakrishnan S, Chowdhary CL, Hakak S, Kaluri R, Praveen Kumar Reddy
M (2020) An ensemble based machine learning model for diabetic retinopathy classification. In: 2020
International Conference on Emerging Trends in Information Technology and Engineering (ic-ETITE), pp
1–6
Ryffel T, Trask A, Dahl M, Wagner B, Mancuso J, Rueckert D, Passerat-Palmbach J (2018) A generic
framework for privacy preserving deep learning. In: PRIVACY PRESERVING MACHINE LEARNING
NeurIPS Workshop
Saba T (2020) Recent advancement in cancer detection using machine learning: Systematic survey of decades,
comparisons and challenges. Journal of Infection and Public Health
Salo F, Injadat M, Nassif AB, Shami A, Essex A (2018) Data mining techniques in intrusion detection sys-
tems: A systematic literature review. IEEE Access 6:56046–56058
Salo F, Injadat M, Moubayed A, Nassif AB, Essex A (2019) Clustering enabled classification using ensemble
feature selection for intrusion detection. In: 2019 International Conference on Computing, Networking
and Communications (ICNC), IEEE, pp 276–281
Sarker A, Ginn R, Nikfarjam A, O’Connor K, Smith K, Jayaraman S, Upadhaya T, Gonzalez G (2015)
Utilizing social media data for pharmacovigilance: a review. Journal of biomedical informatics 54:202–
212
Schiffrin A, Belmonte M (1982) Multiple daily self-glucose monitoring: its essential role in long-term glucose
control in insulin-dependent diabetic patients treated with pump and multiple subcutaneous injections.
Diabetes Care 5(5):479–484
Schiffrin A, Mihic M, Leibel BS, Albisser AM (1985) Computer-assisted insulin dosage adjustment. Diabetes
Care 8(6):545–552
Shen L, Margolies LR, Rothstein JH, Fluder E, McBride R, Sieh W (2019) Deep learning to improve breast
cancer detection on screening mammography. Scientific reports 9(1):1–12
Shi C, Pun CM (2018) Multi-scale hierarchical recurrent neural networks for hyperspectral image classifica-
tion. Neurocomputing 294:82 – 93
Sloane R, Osanlou O, Lewis D, Bollegala D, Maskell S, Pirmohamed M (2015) Social media and phar-
macovigilance: a review of the opportunities and challenges. British journal of clinical pharmacology
80(4):910–920
Sneyers E, De Witte K (2017) The interaction between dropout, graduation rates and quality ratings in uni-
versities. Journal of the Operational Research Society 68(4):416–430
Sommer R, Paxson V (2010) Outside the closed world: On using machine learning for network intrusion
detection. In: 2010 IEEE symposium on security and privacy, IEEE, pp 305–316
Stieglitz S, Dang-Xuan L (2013) Social media and political communication: a social media analytics frame-
work. Social network analysis and mining 3(4):1277–1291
Symeonidis P, Malakoudis D (2016) Moocrec. com: Massive open online courses recommender system. In:
RecSys Posters
Troussas C, Chrysafiadi K, Virvou M (2018) Machine learning and fuzzy logic techniques for personalized
tutoring of foreign languages. In: International Conference on Artificial Intelligence in Education, Springer
International Publishing, pp 358–362
46 MohammadNoor Injadat et al.
Truong HM (2016) Integrating learning styles and adaptive e-learning system: Current developments, prob-
lems and opportunities. Computers in human behavior 55:1185–1193
Tsai M, Wang Y, Kwak M, Rigole N (2019) A machine learning based strategy for election result prediction.
In: 2019 International Conference on Computational Science and Computational Intelligence (CSCI), pp
1408–1410
Ullmann TD (2019) Automated analysis of reflection in writing: Validating machine learning approaches.
International Journal of Artificial Intelligence in Education 29(2):217–257
Van Der Aalst W (2016) Data science in action. In: Process mining, Springer, pp 3–23
Vidyasagar M (2015) Identifying predictive features in drug response using machine learning: opportunities
and challenges. Annual review of pharmacology and toxicology 55:15–34
Vormayr G, Zseby T, Fabini J (2017) Botnet communication patterns. IEEE Communications Surveys Tuto-
rials 19(4):2768–2796
Wang H, Gu J, Wang S (2017) An effective intrusion detection framework based on svm with feature aug-
mentation. Knowledge-Based Systems 136:130–139
Wang J, Yang Y, Mao J, Huang Z, Huang C, Xu W (2016) Cnn-rnn: A unified framework for multi-label
image classification. In: 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp
2285–2294, DOI 10.1109/CVPR.2016.251
Wang P, Wu L, Aslam B, Zou CC (2015) Analysis of Peer-to-Peer botnet attacks and defenses. In: Propagation
phenomena in real world networks, Springer, pp 183–214
Webank’s, AI (2019) Federated AI technology enabler
Wen Y, Li W, Roth H, Dogra P (2019) Federated Learning powered by NVIDIA Clara. Available at:
https://developer.nvidia.com/blog/federated-learning-clara/
Wilson RA, Keil FC (2001) The MIT encyclopedia of the cognitive sciences. MIT press
Wu Q, Zhao W (2017) Small-cell lung cancer detection using a supervised machine learning algorithm. In:
2017 International Symposium on Computer Science and Intelligent Controls (ISCSIC), pp 88–91
Xia F, Shukla M, Brettin T, Garcia-Cardona C, Cohn J, Allen JE, Maslov S, Holbeck SL, Doroshow JH,
Evrard YA, et al. (2018) Predicting tumor cell line response to drug pairs with deep learning. BMC bioin-
formatics 19(18):71–79
Xiao C, Choi E, Sun J (2018) Opportunities and challenges in developing deep learning models using elec-
tronic health records data: a systematic review. Journal of the American Medical Informatics Association
25(10):1419–1428
Xu L, Kinkyo T, Hamori S (2018) Predicting currency crises: A novel approach combining random forests
and wavelet transform. Journal of Risk and Financial Management 11(4):86
Xu R, He M (2020) Application of deep learning neural network in online supply chain financial credit
risk assessment. In: 2020 International Conference on Computer Information and Big Data Applications
(CIBDA), pp 224–232
Yang L, Shami A (2020) On hyperparameter optimization of machine learning algorithms: Theory
and practice. Neurocomputing DOI https://doi.org/10.1016/j.neucom.2020.07.061, URL http://www.
sciencedirect.com/science/article/pii/S0925231220311693
Yang L, Moubayed A, Hamieh I, Shami A (2019a) Tree-based intelligent intrusion detection system in inter-
net of vehicles. In: 2019 IEEE Global Communications Conference (GLOBECOM)
Yang Q, Liu Y, Chen T, Tong Y (2019b) Federated machine learning: Concept and applications. ACM Trans-
actions on Intelligent Systems and Technology (TIST) 10(2):1–19
Zaidi R, Tanveer S (2017) Reviewing anatomy of botnets and botnet detection techniques. International Jour-
nal of Advanced Research in Computer Science 8(5)
Zeng Y, Qiu M, Zhu D, Xue Z, Xiong J, Liu M (2019) Deepvcm: A deep learning based intrusion detection
method in vanet. In: 2019 IEEE 5th Intl Conference on Big Data Security on Cloud (BigDataSecurity),
IEEE Intl Conference on High Performance and Smart Computing, (HPSC) and IEEE Intl Conference on
Intelligent Data and Security (IDS), pp 288–293
Zhang H, Huang T, Lv Z, Liu S, Zhou Z (2018a) Mcrs: A course recommendation system for moocs. Multi-
media Tools and Applications 77(6):7051–7069
Zhang T, Zhang W, Wei X, Haijing H (2018b) Multiple instance learning for credit risk assessment with
transaction data. Knowledge-Based Systems 161:65–77
Zhang W, Zhong J, Yang S, Gao Z, Hu J, Chen Y, Yi Z (2019) Automated identification and grading system
of diabetic retinopathy using deep neural networks. Knowledge-Based Systems 175:12–25
Zhu Y, Zhou L, Xie C, Wang GJ, Nguyen TV (2019) Forecasting smes’ credit risk in supply chain finance with
an enhanced hybrid ensemble machine learning approach. International Journal of Production Economics
211:22–33
Zisser H, Robinson L, Bevier W, Dassau E, Ellingsen C, Doyle III FJ, Jovanovic L (2008) Bolus calculator:
a review of four smart insulin pumps. Diabetes technology & therapeutics 10(6):441–444
