Content uploaded by Mehmet Kokoç
Author content
All content in this area was uploaded by Mehmet Kokoç on Feb 01, 2021
Content may be subject to copyright.
Kokoç, M., Akçapınar, G., & Hasnine, M. N. (2021). Unfolding Students’ Online Assignment Submission Behavioral
Patterns using Temporal Learning Analytics. Educational Technology & Society, 24 (1), 223-235.
223
ISSN 1436-4522 (online) and 1176-3647 (print). This article of the journal of Educational Technology & Society is available under Creative Commons CC-BY-NC-ND
3.0 license (https://creativecommons.org/licenses/by-nc-nd/3.0/). For further queries, please contact Journal Editors at ets.editors@gmail.com.
Unfolding Students’ Online Assignment Submission Behavioral Patterns
using Temporal Learning Analytics
Mehmet Kokoç1, Gökhan Akçapınar2* and Mohammad Nehal Hasnine3
1School of Applied Sciences, Trabzon University, Turkey // 2Faculty of Education, Hacettepe University, Turkey
// 3Research Center for Computing and Multimedia Studies, Hosei University, Japan //
kokoc@trabzon.edu.tr // gokhana@hacettepe.edu.tr // nehal.hasnine.79@hosei.ac.jp
*Corresponding author
ABSTRACT: This study analyzed students’ online assignment submission behaviors from the perspectives of
temporal learning analytics. This study aimed to model the time-dependent changes in the assignment
submission behavior of university students by employing various machine learning methods. Precisely,
clustering, Markov Chains, and association rule mining analysis were used to analyze students’ assignment
submission behaviors in an online learning environment. The results revealed that students displayed similar
patterns in terms of assignment submission behavior. Moreover, it was observed that students’ assignment
submission behavior did not change much across the semester. When these results are analyzed together with the
students’ academic performance at the end of the semester, it was observed that students’ end-of-term academic
performance can be predicted from their assignment submission behaviors at the beginning of the semester. Our
results, within the scope of precision education, can be used to diagnose and predict students who are not going
to submit the next assignments as the semester progresses as well as students who are going to fail at the end of
the semester. Therefore, learning analytics interventions can be designed based on these results to prevent
possible academic failures. Furthermore, the findings of the study are discussed considering the development of
early-warning intervention systems for at-risk students and precision education.
Keywords: Precision education, Temporal learning analytics, Educational data mining, Assignment submission
behavior, Learning performance
1. Introduction
A deeper understanding of online learning experiences is required for learning designers and researchers. Studies
on theory and practice on how students learn individually or in groups in online environments by analyzing
students’ trace data have increased in recent years (Yang et al., 2020). In the last decade, learning analytics (LA)
studies employing machine learning methods have been carried out to gain actionable insights such as at-risk
students’ detection, learning outcome assessment, and drop-out detection for improving the teaching quality and
learning process. Precision education is known to be a relatively new discipline in higher education that uses the
core philosophies of LA and data-driven methods. Precision education is, as addressed by (Yang, 2019), a new
challenge for conventional LA, machine learning, and artificial intelligence for solving critical aspects in online
education such as spotting at-risk, drop-out, low-engaged students as early as possible by analyzing online
learning behaviors (for instance, assignment submission pattern, and engagement with learning materials).
Precision education contributes towards maximizing students’ online learning experiences and value proposition,
and therefore, it uses data from the latest learning technology and integrates student support processes to ensure
the highest quality teaching (Wilson & İsmaili, 2019). One of the goals of precision education is to predict
students’ learning performance by analyzing their online learning behaviors and providing timely intervention
for supporting their learning process (Lu et al., 2018). Furthermore, precision education can be leveraged to
uncover various critical aspects of education including behavioral, cognitive and emotional.
While precision education emphasizes employing artificial intelligence and other data-driven methods on large-
scale datasets collected from technology-enhanced learning environments (i.e., learning management systems,
digital textbooks), data about assignment submission behavior can be explored more within the scope of
precision education. Students’ online assignment submission behavior is a meaningful part of online learning
experiences (Akçapınar & Kokoç, 2020) and has a relationship with procrastination (Yang et al., 2020).
Students’ online assignment submission behavior could reveal information about learning behavior such as how
students’ behavioral patterns of online assignment submission change over time or relationships between
students’ online assignment submission behaviors and their learning performance. These insights on learning
behavior are crucial for teachers to monitor their students’ learning progress, particularly to spot at-risk or
inattentive students as early as possible. Therefore, modeling students’ online learning behaviors hidden in the
learning traces is an important LA contribution for precision education. Thus, modeling students’ online
assignment submission behavior using temporal analysis techniques can provide important insights into the
224
online learning process and help teachers to plan timely interventions for procrastinators and/or at-risk students
for precision education. Furthermore, the temporal aspect of online assignment submission behavior in precision
education has much to offer in diagnosing students’ learning behavior, however, not been explored much.
As of now, much effort has given to explore learning behavior patterns and predict learning performances based
on interaction data, far too little attention has been paid to analyzing temporal and sequential aspects of trace
data of students (Chen, Knight, Wise, 2018; Olsen, Sharma, Rummel, & Aleven, 2020). Several studies have
focused mainly on aggregated data (e.g., the total number of events) without considering temporal aspects of
online learning behaviors (Juhaňák, Zounek, & Rohlíková, 2019). To the best of our knowledge, a limited
number of studies detect patterns in students’ online assignment submission behaviors using temporal analysis
techniques. Considering the importance of temporal analytics in precision education for diagnosing students’
learning, behavioral patterns, and learning performance prediction, this study explored students’ online
assignment submission behavior patterns by using clustering, Markov Chains, and association rule mining
analysis. With these analyses, this study aimed to contribute precision education literature to investigate whether
these patterns can be used to diagnose and predict at-risk students (e.g., who are not going to submit next
assignment and low-performers) as early as possible. Our study addressed the following research questions:
RQ1. What are the students’ behavioral patterns of online assignment submission?
RQ2. How do students’ behavioral patterns of online assignment submission change over time?
RQ3. What are the association rules between students’ online assignment submission behaviors and their
learning performance that can be used to predict at-risk students as early as possible?
This paper aims to employ educational data mining methods for precision education to uncover a core focus of
precision education, namely, understanding learning behavior while the semester progresses. More precisely, this
paper aimed to diagnose and predict at-risk students based on their online assignment submission behavior over
time using temporal LA. Students’ online submission behavioral data were collected from Moodle and analyzed
with regards to- how students’ assignment submission behavior changes over the period of time, finds
association between their assignment submission and final score, and analyzes the factors affecting students’
learning performance at the end of the semester; and visualizes students’ assignment submission patterns so that
the teacher can get an early insight about the students. Thus, the predictive models and the findings of this study
contribute to the core of precision analytics.
2. Background and literature review
2.1. Precision education
Employing artificial intelligence and machine learning techniques in education and psychology has led to
significant developments in related fields such as educational intelligence, self-regulated learning, and precision
education. Depending on the developments in information and communication technologies, a paradigm shift in
learning and teaching has occurred and new pedagogical models have emerged. One of the new educational
models considering personalized learning is precision education. Precision education can be defined as a new
challenge of applying artificial intelligence, machine learning, and LA for improving teaching quality and
learning performance (Yang, 2019). Precision education aims to analyze educational and learner data, predict
students’ performance and provide timely interventions based on learner profiles for enhancing learning (Lu et
al., 2018). For effective learning design in precision education, LA has contributed not only to the dashboards
and intervention tools but also as the conceptual frameworks guiding research experiences.
The ultimate goals of using LA are to increase student success and improve students’ online learning experience
(Pardo & Dawson, 2016). Studies in LA and precision education literature have provided new findings based on
multimodal data and actionable knowledge to increase the learning/teaching context’s effectiveness. There have
been several attempts (e.g., Azcona, Hsiao, & Smeaton, 2019; Tsai et al., 2020) to explore students’ interaction
and behavioral patterns, to predict students’ learning performance based on their online learning behaviors, to
develop early-warning systems for at-risk students, to support students and teachers decision-making processes,
and to investigate effects of interventions and LA dashboards. The results of the aforementioned studies indicate
that LA provides important clues about students’ online learning experiences and LA tools offer personalized
recommendations to students by visualizing and analyzing their trace data to optimize and improve learning. It is
clear that LA and employing educational data mining methods in educational studies contributes to our
understanding of learning.
225
While many studies have been carried out on profiling learners and prediction learning performances based on
interaction data, less attention has been paid to analyzing temporal and sequential aspects of trace data of
students (Chen, Knight, & Wise, 2018; Juhaňák, Zounek, & Rohlíková, 2019). Rather than modeling the
frequency of clicks and interaction of students in an online learning environment, students’ learning paths need
to be modeled based on time and probability (Cerezo, Sánchez-Santillán, Paule-Ruiz, & Núñez, 2016). Thus,
there is an important gap in the relevant field in terms of behavior modeling. To overcome this gap, event logs
reflecting students’ learning experiences have been modeled using temporal analysis and temporal LA approach
(Knight, Wise, & Chen, 2017). The following section is about temporal LA and its implementation in the
educational context. In precision education, the diagnosis of online learning behavior patterns for using
predictive student modeling is vital to provide students real-time intervention.
2.2. Temporal LA and its role in the educational context
Literature in the educational contexts indicates that both individual and collaborative learning do not happen in
one moment (Knight, Wise, & Chen, 2017). In general, learning happens over a period, which is referred to as a
process. Temporal characteristics of students’ learning data contain valuable insights about the time period or
process of occurrence of particular events (Mahzoon et al., 2018). Thus, analyzing time-related data rather than
just frequencies gives more information about the learning process (Knight, Wise, & Chen, 2017). The temporal
analysis of students’ learning data provides a more in-depth insight into individual and collaborative learning
processes (Nguyen, Huptych, & Rienties, 2018; Olsen et al., 2020). What makes temporal analysis vital in online
and blended learning is that modeling transitions between different students’ actions considering temporal
changes enhance our understanding of online learning behavioral patterns. Also, temporal analysis supports a
more robust prediction model of students’ learning performance to make timely interventions for precision
education.
In the temporal analysis, various techniques are employed for modeling students’ behaviors extracted from their
trace data include process mining, sequential pattern mining, Markov chains, and hidden Markov models. While
process mining discovers a process model from the students’ activity sequences, sequential pattern mining finds
the most frequent patterns through a range of action sequences. Markov chains aggregates sequences of students’
actions into transition models and hidden Markov models have been used for discovering students’ behavioral
patterns considering transitions over time (Boroujeni & Dillenbourg, 2019). There is a significant difference
between time-series analysis and temporal LA. While time-series analysis typically looks for recurring patterns
within a time period for numeric features (Mahzoon et al., 2018), temporal analytics methods help researchers
analyze dynamic student data and mode student behaviors over time at different levels of granularity.
There is an increasing trend of temporal analytics methods being used to diagnose students’ online learning
behavior patterns and predict their learning performance based on temporal data for planning timely
interventions (Cheng et al., 2017; Juhaňák, Zounek, & Rohlíková, 2019; Matcha et al., 2019). Previous studies
have shown that temporal analytics is beneficial to predict students’ learning performance (Papamitsiou &
Economides, 2014), to diagnose of learning patterns and behaviors (Boroujeni & Dillenbourg, 2019), to identify
at-risk learners (Mahzoon et al., 2018), to detect learning tactics and strategies (Matcha et al., 2019) and to
explore the relationship between students’ timing of engagement and learning design (Nguyen, Huptych, &
Rienties, 2018). While the importance of analyzing students’ temporal trace data in online and blended learning
has great potential in improving educational practice, applying temporal analytics to student data is less explored
in educational research (Chen, Knight, & Wise, 2018; Knight, Wise, & Chen, 2017). To date, the temporal
analysis of trace data has been mostly employed in modeling students’ online behaviors in the LA field
(Juhaňák, Zounek, & Rohlíková, 2019). These studies highlight the critical role of temporal analysis of trace data
in diagnosing online learning behaviors and predicting students’ further actions. Although temporal analysis has
been used to unlock students’ online learning behaviors such as quiz-taking, content navigation, e-book reading,
and video viewing, few studies have paid attention to exploring online assignment submission behavior patterns.
Therefore, in our study, we intended to use the temporal LA method to model students’ online learning behavior
patterns, specifically students’ trace data while engaging in online assignment activities.
2.3. Online assignment submission behaviors
There is an increasing demand for online assignments to assess the learning process and evaluate learning
performance. Submission of online assignments is one of the most performed online learning activities by
students (Cerezo et al., 2016). In addition, assignment activity is a commonly used LMS component in blended
226
learning environments and fully online courses (Azcona, Hsiao, & Smeaton, 2019). Moreover, several studies
have shown that number of submitted online assignments, assignment scores, and interaction with assignments
are predictors of students’ learning performances (Lu et al., 2018; Zacharis, 2015). According to a study that
modeled LMS-generated interaction data, students’ interaction with assignments and learning tasks are vital
parts of their learning experiences (Kokoç & Altun, 2019). Since online assignments play a meaningful role both
in evaluating to what extent students understand the course subjects and practicing a course topic (Tila & Levy,
2020), online assignment submission behavior can have crucial consequences for learning process assessment.
Thus, the diagnosis of students’ online assignment submission behaviors has been the subject of much attention
in the literature. Previous studies indicated that students who uploaded their assignments previous to the
submission deadline had been better online learning experiences and higher course performance (Akçapınar &
Kokoç, 2020; Paule-Ruiz, Riestra-González, Sánchez-Santillán, & Pérez-Pérez, 2015).
One of the key educational aspects that makes online assignment submission times vital for precision education
is the early identification of students with procrastination tendencies (Yang et al., 2020). Students’ online
assignment submission times have been added to the LA indicators as a proxy measure of academic
procrastination for identifying students at risk of failure (Cormack, Eagle, & Davies, 2020). For example, Yang
et al. (2020) predicted students’ academic performance through submission pattern data reflecting their
procrastination behaviors with an accuracy of 97%. Additionally, previous studies showed that delaying online
assignment submission as a procrastination behavior resulted in lower grades (Cerezo, Esteban, Sánchez-
Santillán, & Núñez, 2017; Cormack, Eagle, & Davies, 2020). This indicates the importance of analyzing online
assignment submission behavior to identify at-risk and procrastinator students for precision education.
Previous studies indicated that the late completion of an online assignment was associated with lower academic
performances and procrastination tendencies (Cormack, Eagle, & Davies, 2020; Yang et al., 2020). Whereas
online assignment submission behavior is essential for the prediction of students’ learning performance and
understanding their online learning experiences, little is still known about it from temporal LA perspectives. To
the best of our knowledge, only one study by Akçapınar and Kokoç (2020) analyzed students’ online assignment
submission behaviors and found that three clusters emerged based on submission behaviors and most of the
students who did not submit the assignment failed in the blended course. Although this study provides valuable
results on the assignment submission behavior process, more LA research is needed to expand our understanding
of online assignment submission behavior in an online and blended learning environment, especially following
temporal analysis and modeling (Azcona, Hsiao, & Smeaton, 2020; Yang et al., 2020). Understanding the
process of students’ online assignment submission behavior can provide important insights into an effective
personalized/adaptive learning environment and help teachers to plan timely interventions for procrastinators
and/or at-risk students for precision education. Thus, our study aims to better understand students’ online
assignment submission transition behaviors by visualizing the patterns and predicting their further assignment
behaviors in a blended learning course. We hope that the study sheds some light on online assignment
submission behavioral patterns and provides actionable knowledge to design timely interventions for improving
learning.
3. Method
In order to answer the research questions, students’ assignment submission data were analyzed using state-of-
the-art educational data mining techniques including clustering, Markov Chains, and association rule mining.
Markov models and clustering and predictive analysis are commonly used in precision education research as
they can generate easy-to-understand models to diagnose and predict at-risk students on time by analyzing their
behavioral data collected from the educational learning environments (Boroujeni & Dillenbourg, 2019). These
methods can also help researchers to understand the transition probabilities of different students’ behaviors that
can be valuable to plan further interventions to prevent possible academic failures. The employed combined
method allows us to obtain interpretable models to understand the students’ assignment submission behavior, its
relation with the academic performance, and changes that happened over time. The data collection and data
analysis processes are explained in detail in the following sections.
3.1. Participants and context
The data were collected from an Operating Systems course offered by a public university in Turkey. A total of
sixty-nine students participated in the study. In this course, Moodle was actively used as a part of the lecture
delivery together with face-to-face lessons. The students' activities in Moodle can be summarized as following
227
the course resources, participating in the discussions, and doing assignments. The assignments included open-
ended questions related to the weekly topics. The purpose of the assignments was to make the students come
prepared for the class. Students are given five-six days before the class to complete the assignments. The starting
time of the class was set as the deadline for the assignment of the last week. During the semester, 10 assignments
were given to the students. In this study, the data related to the assignment given to the students in the 4th, 6th, 8th,
and 10th week were analyzed. These assignments are chosen because they are directly related to course
objectives. The instructor prepares questions in quizzes to promote students’ use of higher-order thinking skills
such as remembering, understanding, applying, analyzing, revising, and creating. An example of a question
related to the disk scheduling topic is given below. In order to answer this question, the students must know how
the disk scheduling algorithms work and apply them to the given context.
Example Question: Let’s take an example where the queue has the following requests with cylinder numbers as
follows: 90, 198, 27, 112, 16, 104, 69, and 60. Assume the head is initially at cylinder 50. Sort incoming requests
according to the SSTF (shortest-seek-time-first) algorithm.
The students submitted their assignments through the Quiz module in Moodle. Among 69 students, 48 students
submitted the first assignment, 57 students submitted the second assignment, 50 students submitted the third
assignment, and 48 students submitted the fourth assignment. The events that students can perform in the
assignment submission process are presented in Table 1. All the activities related to these events were logged in
Moodle’s database with a time stamp.
Table 1. Activities that the students can perform in the assignment submission process
Event
Description
Assignment viewed
The student viewed the assignment module, saw the assignment description, but
did not open the questions.
Attempt started
This is only the case when the student views the assignment for the first time, and
this does not happen again on subsequent visits.
Question viewed
The student’s displaying each question in the assignment is logged in this way.
Displaying the question also means recording the text in the answer field.
Assignment submitted
This happens when the student completes the assignment. The student can submit
the assignment once and then cannot change the answers.
Question reviewed
If the student displays the assignment after the deadline, it will be labeled as a
review. At this stage, the student can view the answer s/he gave or see the grade if
the assignment is graded.
Within the scope of RQ3, the final grades of the students for the Operating Systems course were considered as
an indicator of academic performance. Students took two written exams (i.e., first in the midterm and second in
the final exam) during the semester. Apart from that, they received assignments regularly in Moodle during the
semester. The students’ final grades were calculated by taking 25% of the midterm exam, 25% of their
assignment scores in Moodle, and 50% of the final exam. The final score was used in the data analysis by
categorizing it as “Passed” and “Failed.” The grades were categorized as “Failed” (n = 30, final score < 50) and
“Passed” (n = 39, final score ≥ 50) considering the indicators in the undergraduate regulations of the university.
3.2. Data pre-processing and feature extraction
A total of 9633 activities of 69 students who submitted their assignments before the deadline are exported from
Moodle’s database. The log sequence for a student can include all the events given in Table 1. Also, Assignment
viewed, Question viewed, and Question reviewed events can take place more than once in a log sequence.
Among the examined records, the shortest log sequence contains only 4 records, while the longest log sequence
consists of 268 records. While an average log consists of 45 records, the median value is 39. An example of a log
sequence consisting of 14 records of a student is as follows: Assignment viewed -> Attempt started -> Question
viewed -> Question viewed -> Question viewed -> Question viewed -> Question viewed -> Question viewed ->
Assignment submitted –> Assignment viewed -> Question reviewed-> Question reviewed-> Question reviewed->
Question reviewed. During the data pre-processing, Moodle log records are processed and features were
extracted for each student. This operation was repeated four times for each assignment. Description of the
extracted features are given in Table 2. These features were selected in the light of existing literature (Akçapınar
& Kokoç, 2020; Cerezo et al., 2017; Stiller & Bachmaier, 2019). For example; time-related features (e.g.,
Duration, Time taken) were selected since previous studies showed that time spent on a task is an important
feature while identifying at-risk students as well as understanding their motivation and competencies in
228
metacognitive learning strategies (Stiller & Bachmaier, 2019). Features related to procrastination behavior (e.g.,
Started on, Completed) were also found to be effective while clustering students based on their assignment
submission behaviors (Akçapınar & Kokoç, 2020) and predicting their academic achievements (Cerezo et al.,
2017).
Table 2. Features used in the study and their descriptions
Feature
Description
Attempt count
The number of time student view the questions.
Duration
The amount of time a student spends on an assignment (in minutes).
Started on
The difference between the date and time the assignment was started and the due date (in
hours).
Completed
The difference between the date and time the assignment was submitted and the due date
(in hours).
Time taken
The amount of time it took the student to start and submit the assignment (in hours).
3.3. Data analysis
The study used cluster analysis to group the students according to similar assignment submission behaviors. As a
temporal analysis, Markov Chains were conducted to model transition behaviors of online assignment
submission, and association rule mining was used to build predictive rules based on the students’ behaviors and
academic performances. Since the contents, question types, and the numbers of the questions are varied in
different assignments, the students’ assignment submission behaviors are clustered independently for each
assignment. To map the clusters in different assignments, each assignment should have the same number of
clusters and features. The clustering process was carried out with categorical data. Hence, all features were
categorized into three levels using the equal interval method. Data analysis and visualizations were performed
using the R data mining tool (R Core Team, 2017). Specifically, cluster analysis was carried out using the K-
Modes algorithm with the help of the R package named klaR. Markov Chains analysis was performed using the
Markov Chain package and the association rule mining analysis was performed using the arules package.
4. Results
4.1. What are the students’ behavioral patterns of online assignment submission? (RQ1)
Within the scope of the second research question, it was investigated whether the students' homework
submission behavior changed over time. For this purpose, students were divided into three clusters for each
assignment independently. The number of clusters determined to be three due to the high interpretability of
having high, medium, and low engaged clusters. Whether the three clusters solution fits the data is validated
visually using the Elbow method. The scaled cluster centers' distributions formed after the cluster analysis are
presented in Figure 1 for each assignment. The cluster centers showed that students displayed similar patterns in
all four assignments. For example, the students in the second cluster in Assignment1 and the students in the first
cluster in Assignment2, the students in the third cluster in Assignment3, and the students in the first cluster in
Assignment4 displayed the same pattern. The prominent features of these students are- they start the assignment
at the last moment (StartedOn), spent less time to complete the assignment (Duration), and the number of
questions displayed (AttemptCount) is less. In other words, the students in these clusters submitted the
assignment, but they gave a minimum effort for the assignment. Similarly, the students in Cluster3 in
Assignment1, the students in Cluster2 in Assignment2, the students in Cluster1 in Assignment3, and the students
in Cluster2 in Assignment4 also displayed a similar behavioral pattern. The prominent features of these students
are- they started the assignment much earlier than the given deadline (StartedOn), spent more time to complete
the assignment (Duration), there is a significant difference between the start and end time of the assignment
(TimeTaken), and the number of question views (AttemptCount) is much higher than the other students.
Although most of the students in these clusters complete their assignment submission on the last day, they start
working on the assignment much earlier than the other students and they make much more effort to complete the
assignment. Finally, it is observed that the students in Cluster1 in Assignment1, in Cluster3 in Assignment2, in
Cluster1 in Assignment3, and in Cluster3 in Assignment4, exhibit similar assignment submission patterns. Like
the students in the first group, these students start their assignment submission near the deadline (StartedOn), but
they spend more time completing the assignment than the first group.
229
Figure 1. Box plots of features in different clusters for each assignment
In further analysis, similar clusters in each assignment were labeled as High, Medium, and Low in order to
analyze students who followed a similar assignment submission pattern. Students who did not submit their
assignments are labeled as None. Regarding this analysis, Cluster3 in Assignment I, Cluster2 in Assignment II,
Cluster1 in Assignment III, and Cluster2 in Assignment IV are mapped to the High group. Cluster1 in
Assignment I, Cluster3 in Assignment II, Cluster2 in Assignment III, and Cluster3 in Assignment IV are mapped
to the Medium group. Cluster2 in Assignment I, Cluster1 in Assignment II, Cluster3 in Assignment III, and
Cluster1 in Assignment IV are mapped to the Low group. Students who did not submit their assignments were
manually assigned to the None group. The distribution of students in each group for all assignments are shown in
Table 3.
Table 3. The number of students in each cluster after mapping
Cluster
Assignment I
Assignment II
Assignment III
Assignment IV
High
14
18
18
19
Medium
18
14
19
11
Low
16
25
13
18
None
21
12
19
21
Total
69
69
69
69
4.2. How do students’ behavioral patterns of online assignment submission change over time? (RQ2)
Within the scope of the second research problem, it was investigated whether the homework submission
behavior of the students changed over time. For this purpose, firstly, the transition between the sets in which the
students took part in different assignments is visualized in Figure 2. As seen in the graph, there are transitions
between High-Medium, Medium-High, Medium-Low, Low-Medium, Low-None, and None-Low states. On the
other hand, it is also noticed that there are limited transitions between High-Low, Low-High, High-None, None-
High, Medium-None, and None-Medium states. Markov Chains analysis was used to analyze the transitions
between different states in more detail. In this way, the student’s probabilities of transition from None, Low,
Medium, or High status in one assignment to None, Low, Medium, or High status in another assignment were
calculated. The values calculated for Assignment1-Assignment2, Assignment2-Assignment3, and Assignment3-
Assignment4 transitions are presented in Figure 3.
230
As stated earlier, we clustered students in High, Medium, Low, and None after mapping their assignment
submission behavior. Hence, the Markov Chain analysis in Figure 3 shows the actual transition probabilities
between the groups across the semester.
Figure 2. The students’ assignment submission behaviors over time
Figure 3. Transition probabilities among different assignments
The arrow between the groups indicates the direction of the transition and the numerical values represent the
probability of the transition between each group. The highest probability of each transition is 1 (that is, 100%).
Our Markov Chains analysis uncovered some important assignment submission behaviors of the students;
therefore, we elaborate four key transitions, namely High-to-None, High-to-Low, None-to-High, and Low-to-
High. For High-to-None transition, the Markov Chains analysis indicates that- students in the High cluster who
submitted Assignment I have the transition probability of 0.07 to be in the None cluster in their Assignment II
submission. This means the High-to-None cluster transition is like this that only 7 out of 100 students will not
submit their Assignment II who belonged to the High cluster in their Assignment I submission. Consequently,
for High-to-Low transition, students in the High cluster who submitted Assignment I will have a 0.21 (i.e., 21
231
students out of 100) transition probability to be in the Low cluster in their Assignment II submission. Similarly,
for the None-to-High transition, the probability is 0.1. This means, only 10 out of 100 students who belonged to
the None cluster in their Assignment I submission will be in the High cluster in their Assignment II submission.
In the case of Low-to-High transition behavior, we found that the transition probability of assignment
submission is 0.07 (7 out of 100 students) between Assignment I’s Low cluster and Assignment II’s High
cluster.
4.3. What are the association rules between students’ online assignment submission behaviors and their
learning performance that can be used to predict at-risk students as early as possible? (RQ3)
RQ3 was answered using Association Rule Mining (ARM) analysis. The rules related to passing and failing the
course were filtered among the found rules. As a result, 20 rules for students who passed the course and 14 rules
for students who failed the course were obtained. The list of rules obtained and Support, Confidence, and Lift
values for each rule are presented in Table 4.
Table 4. The list of the association rules extracted
No
LHS
RHS
Support
Confidence
Lift
1
{Assg-IV-Medium}
=>
{Passed}
0.16
1.00
1.77
2
{Assg-II-High,Assg-IV-High}
=>
{Passed}
0.16
1.00
1.77
3
{Assg-III-High,Assg-IV-High}
=>
{Passed}
0.14
1.00
1.77
4
{Assg-I-High,Assg-II-High}
=>
{Passed}
0.13
1.00
1.77
5
{Assg-II-High,Assg-III-High}
=>
{Passed}
0.13
1.00
1.77
6
{Assg-I-High,Assg-III-High}
=>
{Passed}
0.12
1.00
1.77
7
{Assg-I-Medium,Assg-III-High}
=>
{Passed}
0.12
1.00
1.77
8
{Assg-I-Medium,Assg-IV-High}
=>
{Passed}
0.10
1.00
1.77
9
{Assg-I-High,Assg-II-High,Assg-IV-High}
=>
{Passed}
0.10
1.00
1.77
10
{Assg-II-High,Assg-III-High,Assg-IV-High}
=>
{Passed}
0.10
1.00
1.77
11
{Assg-IV-High}
=>
{Passed}
0.26
0.95
1.68
12
{Assg-III-High}
=>
{Passed}
0.25
0.94
1.67
13
{Assg-II-High}
=>
{Passed}
0.25
0.94
1.67
14
{Assg-I-High}
=>
{Passed}
0.19
0.93
1.64
15
{Assg-I-High,Assg-IV-High}
=>
{Passed}
0.14
0.91
1.61
16
{Assg-I-Medium}
=>
{Passed}
0.22
0.83
1.47
17
{Assg-II-Medium}
=>
{Passed}
0.13
0.64
1.14
18
{Assg-III-Medium}
=>
{Passed}
0.16
0.58
1.02
19
{Assg-I-Low}
=>
{Passed}
0.13
0.56
1.00
20
{Assg-III-Low}
=>
{Passed}
0.10
0.54
0.95
21
{Assg-II-None,Assg-IV-None}
=>
{Failed}
0.13
1.00
2.30
22
{Assg-I-None,Assg-III-None,Assg-IV-None}
=>
{Failed}
0.13
1.00
2.30
23
{Assg-I-None,Assg-II-None}
=>
{Failed}
0.10
1.00
2.30
24
{Assg-II-Low,Assg-IV-None}
=>
{Failed}
0.10
1.00
2.30
25
{Assg-I-None,Assg-IV-None}
=>
{Failed}
0.20
0.93
2.15
26
{Assg-I-None,Assg-III-None}
=>
{Failed}
0.17
0.92
2.12
27
{Assg-II-None}
=>
{Failed}
0.16
0.92
2.11
28
{Assg-III-None,Assg-IV-None}
=>
{Failed}
0.16
0.92
2.11
29
{Assg-IV-None}
=>
{Failed}
0.28
0.90
2.08
30
{Assg-I-None}
=>
{Failed}
0.28
0.90
2.08
31
{Assg-I-None,Assg-II-Low}
=>
{Failed}
0.12
0.89
2.04
32
{Assg-III-None}
=>
{Failed}
0.22
0.79
1.82
33
{Assg-IV-Low}
=>
{Failed}
0.14
0.56
1.28
34
{Assg-II-Low}
=>
{Failed}
0.19
0.52
1.20
Rule 1 can be interpreted as- students belonging to the Medium cluster who submitted Assignment 4 on time will
pass at the end of the semester. The confidence of this rule is found to be high (Confidence = 1.0, Support =
0.16, Lift = 1.77). Rule 2 also has a high confidence rate as equal (Confidence = 1.0, Support = 0.16, Lift = 1.77)
as Rule 1, where it is established that- students in the High cluster who submitted both Assignment 2 and
Assignment 4 are likely to pass at the end of the semester. Rules 3 to 10 generated by the association rule mining
analysis have the same confidence (Confidence = 1.0) and lift (Lift = 1.77); however, the support values vary.
232
Rules 11 to 20 that represent the rules for the students who passed at the term-end differ much concerning each
rule’s confidence, support, and lift.
Rule 21 to 34 are for those students who are likely to fail at the end of the semester. For instance, Rule 21
suggests that- the students in the None cluster who had not submitted Assignment 2 and Assignment 4 are likely
to fail in this course. Here, the confidence of our analysis is high (Confidence = 1.0) which means all the
students who are following this pattern failed the course. The rule 22, 23, and 24 for the failed students have
equal confidence (Confidence = 1.0) as Rule 21 revealed that- those students had not submitted Assignment 1-2
& 4, Assignment 1 & 2, and Assignment 2 & 4, respectively.
5. Discussion, conclusion, and limitations
Precision education aims to use artificial intelligence, LA, data analytics, text analytics, image analytics, and
machine learning methods to solve complex educational problems that are yet to uncover in higher education.
Along with LA, precision education also improves teaching quality and learning performance by identifying
inattentive students in the classroom, at-risk students, potential drop-outs, and predicting final scores. By doing
this, precision education aims to assist teachers in re-designing pedagogy, provide special care to those students
in need, and provide timely feedback. A student’s assignment submission is a complex aspect that has always
been crucial for teachers to understand in order to provide timely feedback. In recent days, students are asked to
submit their assignments using online platforms such as Moodle, Blackboard, and Google classroom. Teachers
often find it difficult to understand how well a given assignment is prepared and submitted while using an online
platform. In addition, it is difficult for the teachers to understand a student’s learning process and assess the
learning outcome just by looking at the logs. Therefore, we need to analyze these logs using precision education
guidelines to reveal more insightful learning patterns such as how a student’s online assignment submission
behavior changes as the semester progress or find the association between students’ online assignment
submission behaviors and their final score. Finding these insightful learning patterns are important for teachers
to provide quality education. To date, studies in precision education primarily emphasized online learning
behaviors such as quiz-taking, content navigation, e-book reading, and video viewing. Therefore, most of the
predictive models in the precision education literature are about identifying at-risk and drop-out using online
interaction data such as reading behavior, content viewing behavior, slide navigation behavior, and related.
However, a few studies have been found that analyzed online assignment submission behavior. In addition, to
analyze the online submission behavior, most of the studies have overlooked the temporality (that is, the
temporal analysis of learning interaction data). As mentioned earlier, temporal LA in precision education can
bring new insightful information from online assignment submission behavioral patterns.
This study is conducted to tackle the abovementioned aspects of precision education. In this study, at first, we
employed cluster analysis to profile students based on their online assignment submission behaviors; after that,
we performed the Markov Chains analysis to investigate whether their patterns of online assignment submission
behaviors change over time; and lastly, we applied the association rule mining method to examine the
relationship between students’ online submission behaviors and their course success. Although numerous studies
use educational data mining methods such as clustering, regression, and classification to diagnose students’
online assignment submission behaviors (Yang et al., 2020), temporal analysis has been rarely employed in
educational research (Olsen et al., 2020). Thus, the study combined exploratory methods and temporal LA to
extract actionable knowledge for learning designers and instructors. Our predictive models contribute to the
precision education literature in terms of a deeper understanding of students’ online assignment submission
behavior’s temporal patterns and establish the relation of these temporal patterns with their learning
performances.
The first research question concerns profiling the students based on their online assignment submission
behaviors. It was revealed that the students were clustered into three groups according to similar assignment
submission behaviors. This result is consistent with Akçapınar and Kokoç (2020) findings, where it was found
that the students’ assignment submission data yielded three different clusters. Our results indicate that most of
the students in cluster low and medium started their assignment submissions just before the due date. This result
is likely to be related to academic procrastination behaviors. Procrastination involves delaying an assignment
submission and learning task as long as possible (Yang et al., 2020). It is implied that most of the students had
high procrastination tendencies based on their assignment submission behaviors. Our results are supported by
previous studies indicating that time-related indicators reflected students’ procrastination behaviors in online
learning (Cerezo et al., 2017; Paule-Ruiz et al., 2015; You, 2016). The clusters based on the students’ behaviors
can be used as input to online learning environments to prevent procrastination behaviors. This predictive model
233
can be applied to detect students’ procrastination behavior from their online assignment submission behavioral
data and inform the course instructor about the group of students using procrastination. Hence, our predictive
model would help the instructor in planning an early intervention for those who are using procrastination
regularly in an assigned learning task or a given assignment.
The second research question showed us whether the student followed the same pattern while submitting their
assignments throughout the term. As a result, we found that the probability of shifting between the High and
Low groups was less than 10%. We yield the conclusion that students in the High group have a low probability
of going to the Low or None group. Likewise, students in the Low or None group during the beginning of the
semester have a relatively low probability of going to the high group as the semester progresses. As a result,
students in the Low and None group are at-risk of failing the course at the end of the semester. Nonetheless,
these results support the idea that using temporal analytics provides exciting possibilities to move towards a new
paradigm of assessment that replaces current point-in-time evaluations of learning states (Molenaar & Wise,
2016).
The third research question examined the relationship between students’ assignment submission behavior and
academic performance. The relationship was modeled using association rule mining. A total of 34 rules were
generated which are related to academic performance. In practice, these rules can be used by the instructors or
system designers to understand students’ assignment submission patterns while the semester is in progress and to
plan necessary interventions to prevent possible academic failures. Regarding the early prediction of students’
end of year academic performance following rules can be used. For example, based on Rule 14 it can be
speculated that if a student belongs to the High interaction group in the first assignment s/he will pass the course
with a probability of 0.93. However, if s/he is in the Low group the probability of passing the course decreases to
0.56 (Rule 19). On the other hand, if s/he does not submit the first assignment (Assg-I-None) s/he will fail the
course with a probability of 0.90 (Rule 30). Predictive models can be developed by using these rules to provide
teachers with actionable insights to support their decision-making processes (Romero & Ventura, 2020). Thus,
these rules can also be used to develop a rule-based intervention engine to prevent at-risk students, which is a
core focus of precision education. The rules found in the study could be used as an input for student models in
LA dashboards and intervention engines. Furthermore, researchers can use the rules to design automatic early
interventions for increasing students’ performance for precision education. Similarly, Tsai et al. (2020)
concluded that the dropout prediction model in their study could provide early warnings and interventions to at-
risk students for achieving precision education. It can be mentioned that identifying at-risk students is a key
concern of precision education. Therefore, a LA intervention is required to help them to change their behaviors.
By using these association rules that we generated to address RQ3, the instructor can spot the at-risk students by
using the data from the first assignment (around the 4th week).
In conclusion, the main contribution of the study is unfolding students’ online assignment submission behavior
using temporal LA. Leveraging online assignment submission behavioral data, this study aims to contribute to
precision education literature in various ways, namely by early detection of procrastination behavior, detection
of at-risk students (students in Low and in None group in Figure 2), and generation of association rules for
building a rule-based intervention for the course teacher. Obtained rules can be used to predict students’ end-of-
term academic performance from their assignment submission behaviors at the beginning of the semester. These
predictive models are primarily for instructors, but students can also get benefited from them. By using the
simple visualizations that have been generated by our predictive models, students can take control of their
assignment submissions. For instance, if a student finds him/herself in a Low or None group in the first few
weeks of the semester, s/he can step-up and quickly submit the assignment. Also, a student can control his/her
procrastination behavior. Students can also find their peers who have similar behavior. The study opens up the
space for future studies as well as the design and development of intervention tools based on temporal features of
online assignment submission behaviors. Moreover, the study asks whether clustering analysis, temporal
analysis, and association rule mining analysis could be used to explore specific patterns of assignment
submission behavior. Our results indicate that temporal analysis can be used to detect the students’ online
assignment submission behavior patterns and transitions between related actions. This study also proves that
combining various analytic methods including clustering, Markov Chains, and association rule mining is useful
for modeling temporal patterns of online assignment submission behaviors. This is a methodological
contribution of the study for further studies in precision education, which provides us with deeper insights into
students’ behavior.
This study has some limitations that need to be discussed. First, the small sample size has decreased the
generalizability of our results. To overcome this, a large-scale study in the future may be conducted in the
context of open online courses. Second, this study used a data-driven approach for temporal analysis of students’
234
behaviors. Apart from online assignment submission behavioral features, other features such as the quality of
assignments, learning achievement, gender, and device students used to complete learning tasks are not
analyzed. In developing predictive models, it is important to analyze LMS data combining with multimodal data
to understand the learning process and predictive studies (Olsen et al., 2020). Thus, in future behavior modeling
studies, researchers may collect different data types from different time periods. Third, the present study did not
compare procrastination tendencies, self-regulation skills, and cognitive differences of the students who have the
same sequential behavioral patterns in the learning process. Therefore, future studies regarding the student
modeling of online assignment submission behaviors in precision education would consider these variables.
References
Akçapınar, G., & Kokoç, M. (2020). Analyzing the relationship between student’s assignment submission behaviors and
course achievement through process mining analysis. Turkish Journal of Computer and Mathematics Education
(TURCOMAT), 11(2), 386-401. doi:10.16949/turkbilmat.711683
Azcona, D., Hsiao, I., & Smeaton, A. F. (2019). Detecting students-at-risk in computer programming classes with learning
analytics from students’ digital footprints. User Modeling and User-Adapted Interaction, 29, 759–788. doi:10.1007/s11257-
019-09234-7
Boroujeni, M. S., & Dillenbourg, P. (2019). Discovery and temporal analysis of MOOC study patterns. Journal of Learning
Analytics, 6(1), 16-33. doi:10.18608/jla.2019.61.2
Cerezo, R., Sánchez-Santillán, M., Paule-Ruiz, M. P., & Núñez, J. C. (2016). Students’ LMS interaction patterns and their
relationship with achievement: A case study in higher education. Computers & Education, 96, 42-54.
doi:10.1016/j.compedu.2016.02.006
Cerezo, R., Esteban, M., Sánchez-Santillán, M., & Núñez, J.C. (2017). Procrastinating behavior in computer-based learning
environments to predict performance: A case study in Moodle. Frontiers in Psychology, 8, 1403.
doi:10.3389/fpsyg.2017.01403
Chen, B., Knight, S., & Wise, A. (2018). Critical issues in designing and implementing temporal analytics. Journal of
Learning Analytics, 5(1), 1-9. doi:10.18608/jla.2018.53.1
Cheng, H. N., Liu, Z., Sun, J., Liu, S., & Yang, Z. (2017). Unfolding online learning behavioral patterns and their temporal
changes of college students in SPOCs. Interactive Learning Environments, 25(2), 176-188.
doi:10.1080/10494820.2016.1276082
Cormack, S. H., Eagle, L. A., & Davies, M. S. (2020) A Large-scale test of the relationship between procrastination and
performance using learning analytics. Assessment & Evaluation in Higher Education, 45(7), 1046-1059.
doi:10.1080/02602938.2019.1705244
Juhaňák, L., Zounek, J., & Rohlíková, L. (2019). Using process mining to analyze students’ quiz-taking behavior patterns in
a learning management system. Computers in Human Behavior, 92, 496-506. doi:10.1016/j.chb.2017.12.015
Knight, S., Wise, A. F., & Chen, B. (2017). Time for change: Why learning analytics needs temporal analysis. Journal of
Learning Analytics, 4(3), 7-17. doi:10.18608/jla.2017.43.2
Kokoç, M., & Altun, A. (2019). Building a learning experience: What do learners’ online interaction data imply? In D. G.
Sampson, D. Ifenthaler, J. M. Spector, P. Isaias, & S. Sergis (Eds.), Learning technologies for transforming teaching,
learning and assessment at large scale (pp. 55-70). New York, NY: Springer. doi:10.1007/978-3-030-15130-0_4
Lu, O., Huang, A., Huang, J., Lin, A., Ogata, H., & Yang, S. J. H. (2018). Applying learning analytics for the early prediction
of students’ academic performance in blended learning. Educational Technology & Society, 21(2), 220-232.
Mahzoon, M. J., Maher, M. L., Eltayeby, O., Dou, W., & Grace, K. (2018). A sequence data model for analyzing temporal
patterns of student data. Journal of Learning Analytics, 5(1), 55-74. doi:10.18608/jla.2018.51.5
Matcha, W., Gašević, D., Uzir, N. A. A., Jovanović, J., & Pardo, A. (2019, March). Analytics of learning strategies:
Associations with academic performance and feedback. In Proceedings of the 9th International Conference on Learning
Analytics & Knowledge (pp. 461-470). New York: ACM. doi:10.1145/3303772.3303787
Molenaar, I., & Wise, A. (2016). Grand challenge problem 12: Assessing student learning through continuous collection and
interpretation of temporal performance data. In J. Ederle, K. Lund, P. Tchounikine, & F. Fischer (Eds.), Grand challenge
problems in technology-enhanced learning II: MOOCs and beyond (pp. 59–61). New York, NY: Springer. doi:10.1007/978-
3-319-12562-6_13
Nguyen, Q., Huptych, M., & Rienties, B. (2018). Using temporal analytics to detect inconsistencies between learning design
and student behaviours. Journal of Learning Analytics, 5(3), 120-135. doi:10.18608/jla.2018.53.8
235
Olsen, J. K., Sharma, K., Rummel, N., & Aleven, V. (2020). Temporal analysis of multimodal data to predict collaborative
learning outcomes. British Journal of Educational Technology, 51(5), 1527-1547. doi:10.1111/bjet.12982
Papamitsiou, Z., & Economides, A. A. (2014). Temporal learning analytics for adaptive assessment. Journal of Learning
Analytics, 1(3), 165-168. doi:10.18608/jla.2014.13.13
Pardo, A., & Dawson, S. (2016). Learning analytics: How can data be used to improve learning practice? In P. Reimann, S.
Bull, M. Kickmeier-Rust, R. Vatrapu, & B. Wasson (Eds.), Measuring and Visualizing Learning in the Information-Rich
Classroom (pp. 41-55). New York, NY: Routledge.
Paule-Ruiz, M. P., Riestra-González, M., Sánchez-Santillán, M., & Pérez-Pérez, J. R. (2015). The procrastination related
indicators in e-learning platforms. Journal of Universal Computer Science, 21(1), 7-22. doi:10.3217/jucs-021-01-0007
R Core Team (2017). R: A Language and environment for statistical computing [Computer software]. Vienna, Austria: R
Foundation for Statistical Computing. Available from https://cran.r-project.org/
Romero, C., & Ventura, S. (2020). Educational data mining and learning analytics: An updated survey. Wiley
Interdisciplinary Reviews: Data Mining and Knowledge Discovery, 10(3), e1355. doi:10.1002/widm.1355
Stiller, K., & Bachmaier, R. (2019). Using study times for identifying types of learners in a distance training for trainee
teachers. Turkish Online Journal of Distance Education, 20(2), 21-45. doi:10.17718/tojde.557728
Tila, D., & Levy, D. (2020). Revising online assignments and the impact on student performance at a community college.
Community College Journal of Research and Practice, 44(3), 163-180. doi:10.1080/10668926.2018.1564089
Tsai, S.-C., Chen, C.-H., Shiao, Y.-T., Ciou, J.-S., & Wu, T.-N. (2020). Precision education with statistical learning and deep
learning: A Case study in Taiwan. International Journal of Educational Technology in Higher Education, 17(1), 12.
doi:10.1186/s41239-020-00186-2
Wilson, M. S., & Ismaili, P. B. (2019). Toward maximizing the student experience and value proposition through precision
education. Business Education Innovation Journal, 11(2), 119-124.
Yang, S. J. H. (2019). Precision education: New challenges for AI in education [conference keynote]. In Proceedings of the
27th International Conference on Computers in Education (ICCE) (pp. XXVII-XXVIII). Kenting, Taiwan: Asia-Pacific
Society for Computers in Education (APSCE).
Yang, Y., Hooshyar, D., Pedaste, M., Wang, M., Huang, Y.-M., & Lim, H. (2020). Prediction of students’ procrastination
behaviour through their submission behavioural pattern in online learning. Journal of Ambient Intelligence and Humanized
Computing. doi:10.1007/s12652-020-02041-8
You, J. W. (2016). Identifying significant indicators using LMS data to predict course achievement in online learning.
Internet and Higher Education, 29, 23-30. doi:10.1016/j.iheduc.2015.11.003
Zacharis, N. Z. (2015). A multivariate approach to predicting student outcomes in web-enabled blended learning courses. The
Internet and Higher Education, 27, 44-53. doi:10.1016/j.iheduc.2015.05.002