Content uploaded by Fatma M. Talaat
Author content
All content in this area was uploaded by Fatma M. Talaat on Dec 21, 2023
Content may be subject to copyright.
ORIGINAL ARTICLE
Toward interpretable credit scoring: integrating explainable artificial
intelligence with deep learning for credit card default prediction
Fatma M. Talaat
1,2,3
•Abdussalam Aljadani
4
•Mahmoud Badawy
5,6
•Mostafa Elhosseini
6,7
Received: 17 July 2023 / Accepted: 3 November 2023
The Author(s), under exclusive licence to Springer-Verlag London Ltd., part of Springer Nature 2023
Abstract
In recent years, the increasing prevalence of credit card usage has raised concerns about accurately predicting and
managing credit card defaults. While machine learning and deep learning methods have shown promising results in default
prediction, the black-box nature of these models often limits their interpretability and practical adoption. This study
presents a new method for predicting credit card default using a combination of deep learning and explainable artificial
intelligence (XAI) techniques. Integrating these methods aims to improve the interpretability of the decision-making
process involved in credit card default prediction. The proposed approach is evaluated using a real-world dataset and
compared to existing state-of-the-art models. Results show that the proposed approach achieves competitive prediction
accuracy while providing meaningful insights into the factors driving credit card default risk. The present investigation
adds to the increasing body of literature on explainable artificial intelligence (AI) in the realm of finance. Besides, it
provides a pragmatic approach to assessing credit risk, balancing precision and comprehensibility. In conclusion, the model
demonstrates strong potential as a credit risk assessment tool, with an accuracy of 0.8350, sensitivity of 0.8823, and
specificity of 0.9879. Among the most important features identified by the model are payment delays and outstanding bill
amounts. This study is a step toward more interpretable and transparent credit scoring models.
Keywords Credit risk assessment Deep learning Explainable artificial intelligence (XAI) Financial risk modeling
Machine learning; SHapley Additive exPlanations (SHAP)
1 Introduction
The banking sector, historically volatile and risk-prone, has
witnessed the rapid adoption of credit cards as a primary
financial product in numerous countries. This innovation
allows individuals to borrow and withdraw cash through
credit, thereby introducing a new dimension of credit risk.
As the popularity of credit cards has surged, so has the
associated credit risk. According to industry reports and
statistical data, the prevalence of credit card usage has
reached remarkable heights. A source from Clearly Pay-
ments [18] reveals an exponential growth in the number of
credit card holders, with millions globally relying on them
for daily transactions. Recent studies further underscore
our motivation. A study by the Federal Reserve Bank of
San Francisco [7] found that in 2021, credit cards
accounted for 28% of all payments, marking the highest
level since the study’s inception in 2016. This indicates a
shift towards credit card payments, especially among
higher-income households. Furthermore, a Forbes Advisor
survey from February 2023 reveals that a mere 9% of
Americans predominantly use cash, with 36% opting for
physical or virtual credit cards.
Credit card defaults are a significant problem for both
borrowers and financial institutions. When borrowers fail
to pay their credit card debts, they can face severe financial
consequences, such as damage to their credit score, high
interest rates, and legal action. For financial institutions,
credit card defaults can result in significant losses and
damage to their reputation, as well as regulatory scrutiny
and potential fines.
One of the main drivers of credit card defaults is
financial hardship, such as job loss or medical expenses.
However, other factors, such as irresponsible borrowing
and lending practices, can contribute to the problem.
Extended author information available on the last page of the article
123
Neural Computing and Applications
https://doi.org/10.1007/s00521-023-09232-2(0123456789().,-volV)(0123456789().,-volV)
Financial institutions use various tools and strategies to
address the issue of credit card defaults, such as credit
scoring models and debt collection agencies. However, the
accuracy and effectiveness of these tools depend on factors
such as data quality and algorithm complexity, as men-
tioned by [12].
A credit card default occurs when a borrower fails to
meet the repayment terms and conditions agreed upon with
their credit card issuer [2]. This typically involves not
making the required minimum monthly payments on the
credit card balance for an extended period, usually around
60 to 90 days or more, depending on the card issuer’s
policy [5].
When a borrower defaults on their credit card, several
consequences may follow:
•Late fees: The card issuer may impose late fees for
each missed payment, which can add up quickly and
further increase the debt.
•Higher interest rates: Some credit card issuers may
increase the interest rate (also known as the penalty
rate) on the account after a default, making it more
expensive to carry a balance.
•Credit score impact: If the borrower doesn’t act, their
credit score will be harmed because the credit bureaus
will be notified of the default. Acquiring credit later on
could become harder, leading to elevated interest rates
on loans and other credit offerings.
•Collections: The credit card issuer may transfer the
debt to a collections agency, which will attempt to
recover the outstanding balance from the borrower.
This can involve frequent calls and letters and may even
lead to legal action.
•Charge-off: If the borrower does not pay their debt, the
issuer may eventually write off the debt as a loss,
known as a charge-off. This will be reported to credit
bureaus and will have a long-lasting, negative impact
on the borrower’s credit history.
Borrowers must be mindful of the ramifications of credit
card default and take necessary precautions to evade such a
situation. These measures may include paying at least the
minimum amount due monthly and engaging with the card
issuer to explore alternate payment arrangements, espe-
cially if they are experiencing financial difficulties.
Accurate prediction models for credit card default risk
are pivotal in maintaining the equilibrium of the financial
ecosystem. Their significance spans from individual
financial health to broader economic stability. Specifically:
•Borrowers’Financial Health: Timely payments pre-
vent severe credit score damage, late fees, and potential
legal actions. Defaults can jeopardize future credit
opportunities and overall financial stability.
•Financial Institutions Revenue and Risk: Credit card
companies depend on interest payments and fees from
card balances. Defaults mean lost revenue and potential
debt write-offs. Accurate models allow for proactive
risk mitigation, such as adjusting credit terms based on
borrower risk profiles.
•Broader Financial Ecosystem: Widespread defaults
can strain the entire financial system, leading to
potential economic downturns. Precise prediction mod-
els act as an early warning, enabling proactive risk
management and safeguarding economic stability.
Machine learning and artificial intelligence have wit-
nessed remarkable advancements in recent years, with
various deep learning techniques demonstrating their
potential to solve complex real-world problems.
Researchers have leveraged deep neural networks to
address diverse challenges, ranging from forecasting real
estate housing prices [22] to identifying eye diseases
through image classification [23] and recognizing chaotic
time series patterns [24]. These pioneering studies under-
score the versatility and effectiveness of deep learning
models in handling intricate and dynamic data.
Numerous researchers have contributed to developing
accurate prediction models for credit card default risk. In a
systematic review, [14] analyzed 76 papers from the past
eight years that employ statistical, machine learning, and
deep learning techniques to tackle issues related to credit
risk. The researchers introduce a new approach for classi-
fying credit risk algorithms driven by machine learning and
evaluate their performance using publicly available data-
sets. They also examine the challenges associated with this
field, including imbalanced data, inconsistent datasets, the
transparency of models, and insufficient use of deep
learning models. In their study, Gao et al. [8] employed a
deep learning algorithm, RESNET, to classify customers
with significant unpaid debts. To address the imbalance in
the training data, they used the GAN technique to create
balanced simulated data samples, which enhanced the
model’s effectiveness. As a result, the classification and
accuracy of negative classes improved dramatically.
Explainable artificial intelligence (XAI) is a potential
solution to address the interpretability challenge in
machine learning and deep learning models. By incorpo-
rating XAI techniques, researchers can develop models that
provide accurate predictions and offer insights into the
underlying factors driving these predictions. This added
transparency and interpretability could help build trust in
the models and facilitate their widespread adoption in real-
world applications. XAI, or Explainable Artificial Intelli-
gence, pertains to the capability of an AI system to offer
coherent and easily understandable justifications for its
judgments and behaviors. This is particularly important in
Neural Computing and Applications
123
areas such as credit card default risk prediction, where the
accuracy and transparency of the model are critical for the
success of the decision-making process, as proposed by [9].
In credit card default, explainable AI can help in several
ways. First, it can help to identify which factors or vari-
ables are most important in determining credit card default
risk. This information can be used to refine the model and
improve its accuracy. Second, it can help identify any
biases or errors in the model that might lead to inaccurate
predictions. Finally, it can help build trust and confidence
in the model by allowing users to understand how it works
and why it makes certain predictions [10].
Several techniques and methods can be used to build
explainable AI systems for credit card default. These
include decision trees, rule-based systems, and model-ag-
nostic approaches such as LIME and SHAP. Each method
has unique advantages and disadvantages and selecting a
particular policy hinges upon the specific demands and
limitations of the given problem, as proposed by [6].
In conclusion, using explainable AI in credit card
default is essential for accurate and reliable models. XAI
can help improve the model’s accuracy, identify biases and
errors, and build trust and confidence in the system by
providing clear and understandable explanations of its
decisions and actions.
The main contributions of the paper are:
•Proposing a novel approach for credit card default
prediction based on deep learning and XAI for
interpretability.
•It is evaluated using a real-world dataset and compared
with existing models, achieving accurate prediction and
meaningful insights.
•Contributes to the explainable AI literature in finance,
offering a practical solution for credit risk assessment.
•Highlights interpretability in credit scoring models for
more transparency and balance between accuracy and
interpretability.
•Suggests that the algorithm can be improved and
extended to address related problems in credit risk
management.
This paper is structured as follows: Sect. 2reviews
various topics, such as credit card default, machine learn-
ing, deep learning techniques, and explainable AI in
finance. Section 3outlines the methodology used in this
study, which includes the proposed approach, data pre-
processing, and evaluation metrics. Section 4 discusses the
results obtained and compares the performance of the
proposed approach with existing models while emphasiz-
ing the value added by the explainable AI component.
Finally, Sect. 5concludes the paper by summarizing the
main findings and offering suggestions for further research.
2 Related work
Credit card default prediction has been widely researched
over the years, with various approaches being proposed and
evaluated. However, recent studies have increasingly
focused on machine learning and deep learning techniques,
given their ability to capture complex patterns and auto-
matically learn relevant features from large-scale datasets.
Chen et al. [4] proposed a prediction model based on
k-means SMOTE and BP neural network (BPnn). In this
model, the data distribution is altered using the k-means
SMOTE algorithm, and the significance of the data features
is determined using a random forest model before being
incorporated into the BP neural network’s initial weights
for prediction. The classification performance of these six
prediction models is compared while five standard machine
learning models—KNN, logistics, SVM, random forest,
and tree—were also built in this research.
Deep learning models, particularly neural networks,
have demonstrated promising results in credit risk assess-
ment. For example, authors in [11] proposed a deep neural
network-based model that outperformed traditional meth-
ods like logistic regression and support vector machines in
predicting credit card defaults. Similarly, a convolutional
neural network (CNN) was employed to model temporal
and spatial features of credit card transaction data, resulting
in improved default prediction accuracy [17].
Despite their advantages, the lack of interpretability in
deep learning models has raised concerns, particularly in
the finance domain. Explainable AI (XAI) has emerged as a
potential solution to address this challenge. Various XAI
techniques, such as local interpretable model-agnostic
explanations (LIME) and SHapley Additive exPlanations
(SHAP), have been explored in the context of credit risk
assessment. For instance, [3] integrated LIME with a gra-
dient-boosting model to enhance the interpretability of
credit card default predictions.
Earlier works on credit card default prediction relied on
traditional statistical methods, such as logistic regression
(LR), decision trees (DT), and support vector machines
(SVM). For example, Yeh and Lien [16] compared these
techniques comprehensively, finding that SVMs outper-
formed LR and DT in predicting defaults. However, these
traditional models have limitations in handling complex,
nonlinear patterns in credit data, motivating the shift
toward machine learning and deep learning approaches.
Table 1illustrates the most frequently employed models
proposed to predict credit card default risks, such as
XGBoost (XGB), support vector machine (SVM), random
forest (RF), and neural networks (NN).
XGBoost is a widely used technique for ensemble
learning, which involves merging multiple decision trees to
Neural Computing and Applications
123
enhance the model’s ability to make accurate predictions.
SVM is a binary classification algorithm that tries to find
the optimal hyperplane to separate the data points into
different classes. RF refers to a type of ensemble learning
approach that utilizes several decision trees and chooses a
random subset of features to construct each tree. Finally,
NN is a deep learning algorithm that uses multiple layers to
extract features from the input data. Despite their high
predictive power, these models lack explainability, which
makes it difficult for end-users to understand why a par-
ticular decision was made.
The recent rise in the use of credit cards has led to
concerns about accurately predicting and managing credit
card defaults. While machine learning and deep learning
methods have shown promising results in default prediction
[1], the black-box nature of these models often limits their
interpretability and practical adoption. Explainable artifi-
cial intelligence (XAI) has emerged as a promising
approach to address this issue by enabling the interpreta-
tion of complex models and identifying factors that con-
tribute to their decisions, as mentioned by [13].
In their study, Ali et al. [20] delve into the crucial realm
of explainable Artificial Intelligence (XAI), a pressing
concern as AI systems proliferate, often with opaque
decision-making processes. Their comprehensive study
explores the need for XAI in the context of trust and
comprehensibility of AI models. This research not only
covers foundational concepts but also delves into recent
advancements in XAI, particularly focusing on supervised
machine learning. The study categorizes XAI techniques
into four axes: data explainability, model explainability,
post hoc explainability, and assessment of explanations,
providing a holistic view of the landscape. Furthermore, it
introduces valuable evaluation metrics, open-source pack-
ages, and datasets. The study underlines explainability’s
legal, user-centric, and application-oriented aspects. With a
thorough examination of 410 key articles published
between January 2016 and October 2022, this study serves
as an indispensable resource for XAI researchers striving to
enhance the trustworthiness of AI models, offering insights
and direction for researchers from diverse disciplines
seeking effective XAI methods.
In their comprehensive review of XAI applications in
finance, [21] emphasizes the growing importance of
Explainable Artificial Intelligence (XAI) in highly regu-
lated domains such as finance, where the adoption of
technological advances like Artificial Intelligence requires
transparency and traceability of decisions. The systematic
literature review screened over 2,000 articles and identified
60 relevant articles, classifying them based on the XAI
methods and goals they employed. Their findings highlight
the extensive research in areas like risk management and
portfolio optimization while underscoring the need for
more attention to anti-money laundering applications. The
review also notes the utilization of transparent models and
post hoc explainability, with a recent preference for the
latter in the finance domain.
Lange et al. [19] introduced an explainable artificial
intelligence (XAI) model designed for predicting credit
default, leveraging a distinctive dataset of unsecured con-
sumer loans from a Norwegian bank. Their approach
combined a LightGBM model with SHAP, enabling the
interpretation of explanatory variables influencing predic-
tions. Notably, the LightGBM model demonstrated supe-
rior performance compared to the bank’s conventional
credit scoring model, a Logistic Regression model. Key
findings unveiled that the most critical explanatory vari-
ables for default prediction in the LightGBM model
encompassed the volatility of the utilized credit balance,
remaining credit as a percentage of total credit, and the
duration of the customer relationship. This study consti-
tutes a noteworthy contribution to applying XAI techniques
in banking, shedding light on their potential to enhance the
interpretability and reliability of advanced AI models.
3 Research gap
Although machine learning and deep learning techniques
have significantly improved credit card default prediction,
their ‘‘black-box’’ nature presents a major challenge
regarding interpretability and transparency. In finance,
stakeholders often require transparent and explainable
decision-making processes to comply with regulatory
requirements and ensure ethical practices. As such, there is
a need to develop models that balance predictive accuracy
with interpretability. The research gap can be summarized
as follows:
•The black-box nature of these models raises inter-
pretability concerns.
•Regulatory requirements and ethical practices demand
transparency.
•Existing literature focuses on integrating XAI with
traditional machine learning.
Table 1 Comparison of performance results for commonly used
models in credit risk assessment
Model Accuracy Sensitivity Specificity
XGB 0.7978 0.8429 0.7057
SVM 0.7916 0.8410 0.6906
RF 0.7928 0.8521 0.6717
NN 0.7531 0.8281 0.6000
Neural Computing and Applications
123
•Deep learning models with superior performance have
not been extensively explored with XAI.
The present study endeavors to fill this void by con-
structing and assessing a novel approach that amalgamates
the predictive capabilities of deep learning methodologies
with the explicable qualities of explainable artificial intel-
ligence (XAI). In doing so, we hope to achieve the
following:
•Balance accuracy and transparency for credit risk
assessment.
•Foster trust among stakeholders and facilitate real-
world adoption of these models.
•Contributing significantly to credit card default predic-
tion and the broader area of explainable AI in finance.
4 CreditNetXAI: integrating explainable AI
with deep learning for credit card default
prediction
This section proposes a novel deep neural network
approach called CreditNetXAI for credit card default
prediction, which integrates deep learning and explainable
AI techniques. CreditNetXAI aims to achieve high pre-
dictive accuracy while providing interpretable and trans-
parent results.
CreditNetXAI approach is composed of four main
phases, as depicted in Fig. 1: (i) data collection and pre-
processing, (ii) model development, (iii) model training
and validation, and (iv) evaluation and interpretation.
i. Data Collection and Preprocessing: this phase
contains the following steps: (i) Collect a real-world
dataset containing credit card user information,
transaction data, and default records. (ii) Divide the
dataset into training and testing sets for model
development and evaluation. (iii) Conduct data
preprocessing, addressing missing values, detecting
and treating outliers, scaling features, and encoding
categorical variables. (iv) Utilize methods for feature
selection, such as recursive feature elimination
(RFE) or LASSO, to determine the most significant
attributes for forecasting credit card defaults.
ii. Model Development: this phase contains the fol-
lowing steps: (i) Develop the CreditNetXAI
approach, which integrates deep learning and
explainable AI techniques for credit card default
prediction. (ii) Implement the deep learning compo-
nent using a neural network architecture tailored to
the task, with multiple layers and activation func-
tions. (iii) To ensure the model can generalize well to
new data, it is important to prevent overfitting during
training. This can be achieved by incorporating
methods such as dropout and early stopping. (iv)
Incorporate the XAI component, which employs a
model-agnostic explanation method (SHAP) to gen-
erate feature importance scores and provide insights
into the factors driving the model’s predictions.
iii. Model Training and Validation: this phase con-
tains the following steps: (i) Train the CreditNetXAI
approach on the preprocessed training data, using
dropout and early stopping techniques to prevent
overfitting. (ii) Carry out hyperparameter tuning
through grid or random search techniques to opti-
mize the model’s performance. (iii) Validate the
trained model using cross-validation to assess its
generalization performance and ensure its
robustness.
iv. Evaluation and interpretation: this phase contains
the following steps: (i) Evaluate the performance of
the CreditNetXAI approach using a combination of
metrics that capture the model’s predictive accuracy
and interpretability aspects. (ii) Accuracy-related
metrics may include area under the receiver operat-
ing characteristic curve (AUROC), F1-score, preci-
sion, recall, and accuracy. (iii) Interpretability-
related metrics may involve the explanation quality
of the XAI component, such as fidelity, stability, and
consistency of the generated feature importance
scores. (iv) Use the XAI component to generate
feature importance scores and provide insights into
the factors driving the model’s predictions, enhanc-
ing the interpretability and transparency of the
CreditNetXAI algorithm.
4.1 Data collection and preprocessing
The data collection and preprocessing phase combines four
main steps as illustrated in Algorithm 1. (i) Step 1: Col-
lecting the dataset; (a) Collecting a real-world dataset
containing credit card user information, transaction data,
and default records. (b) Store the dataset in a suitable for-
mat (e.g., CSV, Excel, SQL database). (ii) Step 2: Split-
ting the dataset; (a) Divide the dataset into training and
testing sets for model development and evaluation. (b) The
split can be random or stratified based on the target variable
(default status). (iii) Step 3: Preprocessing the dataset;
(a) Handle missing values by imputing or removing them.
(b) Detect and treat outliers in the data to prevent them
from affecting the model’s performance. (c) To ensure they
have similar ranges, scale the numerical features using
min-max scaling or standardization techniques. (d) Encode
categorical variables using one-hot or label encoding
techniques to convert them to numerical values that the
Neural Computing and Applications
123
model can understand. (iv) Step 4: Feature selection;
(a) Use feature selection techniques like recursive feature
elimination (RFE) or LASSO to identify the most relevant
features for credit card default prediction. (b) RFE can be
performed using a wrapper method, where the model is
trained and evaluated on different subsets of features, and
the best-performing subset is selected. (c) LASSO can be
used as a regularization technique to shrink the coefficients
of irrelevant features to zero, thus eliminating them from
the model.
Fig. 1 The architecture of CreditNetXAI
Neural Computing and Applications
123
Algorithm 1 Data collection and preprocessing algorithm
4.2 Model development
The model development phase combines four main steps as
illustrated in Algorithm 2. (i) Step 1: Develop the Cred-
itNetXAI algorithm; (a) Define the problem statement and
objective of the algorithm. (b) Determine the data
requirements, including data sources, preprocessing, and
cleaning. (c) Choose appropriate performance metrics,
such as accuracy, precision, recall, and F1 score, to eval-
uate the model’s performance. (ii) Step 2: Implement the
deep learning component; (a) Define the neural network
architecture, including the number of layers, nodes per
layer, and activation functions. (b) Define the loss function,
optimizer, and learning rate. (c) Split the data into training,
validation, and test sets. (d) Train the model using the
training set and monitor the validation loss to prevent
overfitting. (e) Tune hyperparameters using the validation
set, such as the number of layers, nodes per layer, and
learning rate. (f) Evaluate the model’s performance on the
test set. Step 3: To guarantee the efficacy of the model
when presented with novel data, it is crucial to circumvent
overfitting during the training phase. This can be accom-
plished by assimilating techniques such as dropout and
early stopping;(a) Implement dropout regularization to
Neural Computing and Applications
123
dropout nodes during training to reduce overfitting ran-
domly. (b) Use early stopping to stop the training process
when the validation loss stops improving to prevent over-
fitting and save computation time. Step 4: Incorporate the
XAI component; (a) Use a model-agnostic explanation
method (SHAP) to generate feature importance scores and
provide insights into the factors driving the model’s pre-
dictions. (b) Use the XAI component to generate expla-
nations for individual predictions and global feature
importance scores for the entire dataset.
Algorithm 2 Model development algorithm
4.3 Model training and validation
The model training and validation phase combines three
main steps as illustrated in Algorithm 3. (i) Step 1: Train
the CreditNetXAI algorithm on the preprocessed
training data; (a) Split the preprocessed training data into
training and validation sets. (b) Initialize the neural net-
work architecture and its hyperparameters. (c) Train the
model on the training set using backpropagation and gra-
dient descent. (d) Use dropout and early stopping tech-
niques to prevent overfitting. (ii) Step 2: Perform
hyperparameter tuning using grid or random search to
optimize the model’s performance; (a) Define a range of
Neural Computing and Applications
123
values for each hyperparameter to be tuned. (b) Perform a
grid or random search over the hyperparameter space.
(c) Evaluate the performance of each model using cross-
validation. (iii) Step 3: Validate the trained model using
cross-validation to assess its generalization performance
and ensure its robustness; (a) Split the preprocessed data
into k-folds. (b) The model should be trained and evaluated
k times, with a different subset of data used for validation
in each iteration while the rest is used for training.
(c) Calculate the average performance metric across the
kiterations.
Algorithm 3 Model training and validation algorithm
Table 2 Features of default of credit card clients data set
Feature Description
ID The ID of each client
LIMIT_BAL Amount of given credit in NT dollars (includes individual and family/supplementary credit
SEX Gender (1 = male, 2 = female)
EDUCATION (1 = graduate school, 2 = university, 3 = high school, 4 = others, 5 = unknown, 6 = unknown)
MARRIAGE Marital status (1 = married, 2 = single, 3 = others)
AGE Age in years
PAY_0 Repayment status in September, 2005 (-1 = pay duly, 1 = payment delay for one month, 2 = payment delay for
two months, …8 = payment delay for eight months, 9 = payment delay for nine months and above)
PAY_2 Repayment status in August 2005 (scale same as above)
PAY_3 Repayment status in July 2005 (scale same as above)
PAY_4 Repayment status in June
PAY_5 Repayment status in May
PAY_6 Repayment status in April
BILL_AMT1 Amount of bill statement in September
BILL_AMT2 Amount of bill statement in August
BILL_AMT3 Amount of bill statement in July
BILL_AMT4 Amount of bill statement in June
BILL_AMT5 Amount of bill statement in May
BILL_AMT6 Amount of bill statement in April
PAY_AMT1 Amount of previous payment in September
PAY_AMT2 Amount of previous payment in August
PAY_AMT3 Amount of previous payment in July
PAY_AMT4 Amount of previous payment in June
PAY_AMT5 Amount of previous payment in May
PAY_AMT6 Amount of previous payment in April
Default.payment.next.month Default payment (1 = yes, 0 = no)
Table 3 A sample of the credit card clients data set: a snapshot of variables and observations
Neural Computing and Applications
123
4.4 Evaluation and interpretation
The evaluation and interpretation phase combines five
main steps as illustrated in Algorithm 4. (i) Step 1: Define
evaluation metrics; (a) Define accuracy-related metrics,
such as AUROC, F1-score, precision, recall, and accuracy.
(b) Define interpretability-related metrics, such as fidelity,
stability, and consistency of the generated feature impor-
tance scores. (ii) Step 2: Split data into training and test
sets; Split the dataset into training and test sets to train and
evaluate the CreditNetXAI algorithm. (iii) Step 3: Train
the CreditNetXAI algorithm; (a) Train the deep learning
component of the CreditNetXAI algorithm on the training
set. (b) Generate feature importance scores using the XAI
component on the training set. (iv) Step 4: Evaluate the
CreditNetXAI algorithm; (a) Evaluate the performance of
the CreditNetXAI algorithm using the evaluation metrics
defined in Step 1. (b) Calculate accuracy-related metrics on
the test set, such as AUROC, F1-score, precision, recall,
and accuracy. (c) Calculate interpretability-related metrics
on the test set, such as fidelity, stability, and consistency of
the generated feature importance scores. (v) Step 5:
Interpret the results; (a) Use the XAI component to
generate feature importance scores on the test set and
provide insights into the factors driving the model’s pre-
dictions. (b) Analyze the feature importance scores and
insights generated by the XAI component to enhance the
interpretability and transparency of the CreditNetXAI
algorithm.
Neural Computing and Applications
123
Algorithm 4 Evaluation and interpretation algorithm
4.5 Methodology explanation
This section comprehensively explains the methodology
employed in the CreditNetXAI algorithm for credit card
default prediction. The methodology comprises several
phases designed to ensure accurate predictions while
maintaining transparency and interpretability.
4.6 Explainable AI techniques
In this section, we delve deeper into the explainable AI
(XAI) techniques employed in our proposed CreditNetXAI
approach and explain how their integration enhances the
model’s interpretability.
CreditNetXAI combines deep learning with XAI tech-
niques, specifically employing the SHAP (SHapley Addi-
tive exPlanations) model-agnostic explanation method. The
objective of integrating XAI is to provide high predictive
accuracy and interpretable and transparent results for credit
card default prediction.
4.6.1 SHAP (SHapley Additive exPlanations)
SHAP is a widely recognized model-agnostic explanation
method rooted in cooperative game theory. It assigns fea-
ture importance scores to individual features in a predic-
tion, quantifying their impact on the prediction’s outcome.
These scores are based on Shapley values, which attribute
contributions to each feature considering all possible
combinations of feature values.
The integration of SHAP enhances the interpretability of
the model in the following ways:
•Feature Importance Scores: SHAP generates feature
importance scores for each prediction, indicating the
magnitude and direction of each feature’s influence on
the prediction. These scores can be used to identify the
critical features that contribute most to the prediction.
•Global Feature Importance: By aggregating feature
importance scores across multiple predictions, SHAP
provides global feature importance scores. These scores
reveal which features, on average, have the most
significant impact on the model’s predictions, enhanc-
ing our understanding of the model’s behavior.
•Individual Explanations: SHAP can provide explana-
tions for individual predictions by attributing the
contribution of each feature to a specific prediction.
This allows stakeholders to understand why a particular
prediction was made and which features drove that
prediction.
Neural Computing and Applications
123
4.7 Choice of deep learning architectures
and hyperparameters
The selection of specific deep learning architectures and
hyperparameters in CreditNetXAI is not arbitrary but based
on a rigorous model development phase. During experi-
mentation, we considered various architectures and
hyperparameter settings to optimize the model’s
performance.
4.8 Reasoning behind choices
•Architecture: We chose the neural network architec-
ture based on its capacity to capture complex patterns
and relationships in credit card default data. The
architecture consists of multiple layers with appropriate
activation functions to learn from the data effectively.
•Hyperparameters: Our choice of hyperparameters,
such as learning rates, dropout rates, and the number of
layers and nodes per layer, was guided by empirical
results and best practices in deep learning. These
parameters were carefully tuned to prevent overfitting
and maximize the model’s generalization performance.
5 Implementation and evaluation
This section introduces the used datasets, the performance
metrics, and the performance evaluation. The computa-
tional setup is:
Hardware:
•CPU: Intel Core i7-8700 K 3.70 GHz.
•GPU: NVIDIA GeForce RTX 2080 Ti.
•RAM: 32 GB.
Software:
•Operating System: Windows 10.
•Python: We used Python as the primary programming
language for our experiments.
•Deep Learning Frameworks: We leveraged popular
deep learning frameworks, including TensorFlow and
Keras, to develop and train our CreditNetXAI model.
•Libraries: Various Python libraries, such as NumPy,
Pandas, Matplotlib, and Scikit-learn, were used for data
preprocessing, analysis, and visualization.
5.1 Default of credit card clients data set
The Default of Credit Card Clients Data Set [15]isa
widely used dataset in the field of credit scoring, which is
used to predict the likelihood of default of credit card
clients. This dataset consists of 30,000 observations of
credit card holders in Taiwan, with 25 features that include
demographic information, credit limit, payment history,
and bill amounts.
The dataset was collected from a financial institution’s
credit card clients in Taiwan and obtained from the UCI
Machine Learning Repository. The data covers the period
from April 2005 to September 2005 and includes infor-
mation on whether or not each credit card holder defaulted
on their payment the following month. Table 2illustrates
the features and their description.
This dataset aims to develop models that can accurately
predict the probability of default for each credit card holder
based on the available features. As a result, financial
institutions can use this information to make more
informed decisions when it comes to extending credit to
potential borrowers. A sample of the Credit Card Clients
Data Set is shown in Table 3.
5.2 Dataset description
In the evaluation phase of our research, we employed the
Default of Credit Card Clients Data Set, sourced from [15].
This dataset is a widely recognized benchmark in the field
of credit scoring. It encompasses 30,000 observations of
credit card holders in Taiwan, featuring 25 distinct fea-
tures, including demographic information, credit limits,
payment history, and bill amounts. We collected this data
from a financial institution’s credit card clients in Taiwan,
spanning the period from April 2005 to September 2005. It
is essential to provide additional details about this dataset:
•Dataset Size: The dataset comprises 30,000 observa-
tions, making it a substantial and representative dataset
for credit card default prediction.
•Class Distribution: The dataset is imbalanced con-
cerning the target variable, which indicates whether a
credit card holder defaulted on their payment the
following month. The class distribution is approxi-
mately 22,120 non-default cases (class 0) and around
7,880 default cases (class 1).
•Data Biases and Limitations: We acknowledge that
real-world datasets, including the one used in this study,
may exhibit inherent biases and limitations. These
biases can stem from factors such as sampling methods,
data collection processes, or the specific population
under study. Therefore, while our model demonstrates
effectiveness on this dataset, it is crucial to exercise
caution when applying it to other contexts or datasets
exhibiting different distributions and biases.
Regarding data preprocessing:
Neural Computing and Applications
123
1. Data Augmentation: We did not perform data
augmentation on this dataset. The dataset was used in
its original form without any synthetic data generation.
2. Handling Missing Values: Missing values, if any,
were treated through established techniques such as
mean imputation, median imputation, or forward-fill/
back-fill, depending on the nature of the missing data
within specific features. However, it’s important to
note that this dataset is well-curated and did not
contain significant missing data.
3. Treatment of Outliers: Outliers were identified using
statistical methods like the interquartile range (IQR)
and were carefully assessed. In cases where outliers
were deemed genuine data points, they were retained in
the dataset as they could indicate real-world scenarios.
However, outliers that were considered erroneous were
either corrected or removed, ensuring data quality and
consistency.
These steps ensured that the dataset used for model
training and evaluation was as representative and reliable
as possible.
5.3 Performance metrics
The performance metrics used in this research paper are: (i)
accuracy: accuracy measures the proportion of correct
predictions among all predictions made by a classifier. It
can be calculated as in Eq. (1). (ii) Sensitivity, also known
as recall or true positive rate, measures the proportion of
actual positive instances correctly identified by the classi-
fier. It can be calculated as in Eq. (2). (iii) Specificity
measures the proportion of actual negative instances cor-
rectly identified by the classifier. It can be calculated as in
Eq. (3).
Accuracy ¼TP þTNðÞ
TP þTN þFP þFNðÞ
ð1Þ
Sensitivity ¼TP
TP þFNðÞ ð2Þ
Specificity ¼TN
TN þFPðÞ ð3Þ
where TP, TN, FP, and FN are the true positive, true
negative, false positive, and false negative counts,
respectively.
5.4 Performance evaluation
The SHAP XAI model is employed to elucidate the vari-
ables influencing the model’s predictions to foster trans-
parency and comprehensibility. Figure 2illustrates the
feature’s importance. Figure 3illustrates the data visual-
ization. Finally, Fig. 4shows the data heatmap.
Table 4and Fig. 5compare the results of the proposed
CreditNetXAI and the previous models.
Table 4highlights that the CreditNetXAI model exhibits
favorable outcomes, with an accuracy of 0.8350, signifying
that it can correctly classify most credit applicants. The
sensitivity value of 0.8823 indicates that the model can
identify the majority of the true positive cases, i.e., credit
applicants who are likely to default on their payments.
Moreover, the specificity of 0.9879 signifies that the model
can accurately identify the majority of the true negative
cases, i.e., credit applicants who are expected to repay their
credit on time.
Figure 5and Table 4provide a comprehensive com-
parison of the CreditNetXAI model’s performance against
several well-known models, including XGBoost (XGB),
support vector machine (SVM), random forest (RF), and
neural network (NN). These models were evaluated on
various performance metrics, including accuracy, sensi-
tivity, and specificity.
The results clearly indicate that the CreditNetXAI model
stands out in terms of performance:
•Accuracy (0.8350): The CreditNetXAI model exhibits
a high level of accuracy, correctly classifying approx-
imately 83.50% of credit card applicants. This indicates
its capability to make accurate predictions and distin-
guish between default and non-default cases.
•Sensitivity (0.8823): Sensitivity, also known as recall
or the true positive rate, is a crucial metric in credit risk
assessment. The CreditNetXAI model excels in this
regard, with a sensitivity value of approximately
88.23%. This means that the model can identify the
majority of true positive cases, effectively recognizing
credit applicants who are likely to default on their
payments.
•Specificity (0.9879): Specificity measures the ability of
the model to identify true negative cases, i.e., applicants
who are expected to repay their credit on time. The
CreditNetXAI model achieves an impressive specificity
value of about 98.79%, indicating its ability to distin-
guish non-default cases accurately.
These results underscore the CreditNetXAI model’s
effectiveness in credit card default prediction. Its robust
performance, particularly in identifying default and non-
default cases, positions it as a valuable tool for financial
institutions and credit assessors.
5.4.1 Interpretability through SHAP
Figures 2,3, and 4demonstrate the interpretability features
of the CreditNetXAI model achieved through SHAP. These
Neural Computing and Applications
123
Fig. 2 Feature importance for
credit card default prediction
using SHAP Model
Fig. 3 Visualization of the credit card clients data set
Neural Computing and Applications
123
visualizations offer insights into the factors that influence
the model’s predictions and enhance transparency and
comprehensibility.
•Feature Importance (Fig. 2): Fig. 2showcases the
feature importance determined by the SHAP model.
This visualization illustrates that PAY_0, PAY_3,
PAY_5, and BILL_AMT2 were the most important
features identified by the neural network.
•Visualization of Data (Fig. 3): This visualization aids
in understanding the underlying data distribution and
relationships among features. It offers a valuable
perspective on how different factors may be correlated
and influence credit card default risk. Among the most
important features identified by the model are payment
delays and outstanding bill amounts.
•Data Heatmap (Fig. 4): The heatmap visually repre-
sents data correlations and can reveal patterns con-
tributing to credit card default. Identifying these
patterns is vital for understanding the factors driving
credit risk.
In conclusion, the CreditNetXAI model excels in pre-
dictive accuracy and distinguishes itself through its inter-
pretability. Using SHAP empowers stakeholders to gain
insights into the model’s decision-making process and
understand why specific applicants are deemed risky or
creditworthy. This transparency is essential in the domain
of credit risk assessment, as it enables informed decision-
Fig. 4 Heatmap of the credit
card clients data set
Table 4 Comparison of CreditNetXAI versus previous models for
credit card default prediction
Model Accuracy Sensitivity Specificity
CreditNetXAI 0.8350 0.8823 0.9879
XGB 0.7978 0.8429 0.7057
SVM 0.7916 0.8410 0.6906
RF 0.7928 0.8521 0.6717
NN 0.7531 0.8281 0.6000
Fig. 5 Performance comparison of CreditNetXAI Versus previous
models for credit card default prediction
Neural Computing and Applications
123
making and ultimately contributes to more responsible
lending practices.
5.5 Model explanations: importance of variables
According to the analysis conducted with XGBoost, factors
like gender, marital status, and the amount paid in the sixth
month are not significant predictors. On the other hand, the
delay in making the first month’s payment, which may
have initially been considered unimportant, appears to have
a notable influence on the outcome. It is worth noting that
outstanding bill amounts are more significant in predicting
credit default than an individual’s age. This suggests that
higher outstanding balances might be a stronger predictor
of payment default than older individuals with lower
incomes.
According to the SVM model, the variables that appear
to be significant are age, sex, education, and marriage. In
addition, payments made in the 5th and 6th months seem
crucial for prediction. The SVM model also places greater
importance on the limit_bal variable than other models.
Based on these findings, it can be concluded that during the
latter part of the year, when payments accumulate, an
individual’s capability to pay off all or most of their out-
standing balances is the primary factor in determining
whether they will default on their payments or not.
The random forest algorithm reveals that payment
delays within the last three months are the most significant
indicators of credit card defaults, which is a surprising
finding. Conversely, the algorithm ranks age as one of the
least important variables in predicting credit card defaults.
This suggests that payment delays in later months, where
unpaid bills accumulate, are strong indicators of default,
regardless of the borrower’s age.
The neural network is the least accurate model among
the five. This can be attributed to its inability to differen-
tiate the variables with the greatest discriminatory power,
which results in more inaccurate predictions.
The CreditNetXAI model was the most effective among
the four models used in the study. This model revealed that
PAY_0, PAY_3, PAY_5, and BILL_AMT2 were the most
important features identified by the neural network.
5.6 Results discussion
The CreditNetXAI model showed promising results with
an accuracy of 0.8350, indicating that the model could
classify the majority of the credit applicants correctly.
Furthermore, the sensitivity of 0.8823 suggested that the
model could identify the majority of the true positive cases,
i.e., the applicants who were likely to default on their credit
payments. In addition, the specificity of 0.9879 indicated
that the model correctly identified the majority of the true
negative cases, i.e., the applicants who were likely to repay
their credit on time. These results suggest that the Cred-
itNetXAI model can be an effective tool for credit risk
assessment. It has high accuracy and can identify positive
and negative cases with high sensitivity and specificity.
The CreditNetXAI model provides insights into the
factors contributing to its predictions, making it a trans-
parent and explainable model. This can be particularly
useful in the context of credit risk assessment, as it can help
credit assessors understand why certain applicants are
deemed risky and make informed decisions based on these
insights. In addition, the proposed algorithm balances
accuracy and interpretability, contributing to the growing
literature on explainable AI in finance.
Overall, the CreditNetXAI model shows promise as an
effective and transparent tool for credit risk assessment.
However, further evaluation and testing may be necessary
to ensure that the model is effective in different scenarios
and for different types of credit applications. In addition,
the proposed algorithm can be further improved and
extended to address related problems in credit risk man-
agement. This study is a step toward more interpretable and
transparent credit scoring models, and the CreditNetXAI
model offers a practical solution for credit risk assessment
that balances accuracy and interpretability.
5.7 Model robustness and limitations
We understand the importance of assessing the robustness
and generalization capability of the proposed Cred-
itNetXAI model when applied to various datasets and real-
world scenarios. To address this concern, we have included
the following information:
Challenges may arise when applying the CreditNetXAI
model to different datasets or real-world scenarios due to
data variability, cultural and regional financial differences,
and evolving financial patterns. The model’s inter-
pretability might differ across contexts, and external eco-
nomic shocks or regulatory changes could influence its
predictions. Additionally, potential biases in the original
training data could lead to skewed results in diverse
demographics, and the model’s complexity might pose
deployment challenges in resource-constrained environ-
ments. The robustness of the CreditNetXAI model, when
subjected to diverse datasets or real-world contexts,
remains a pertinent concern. Variabilities in data, regional
financial nuances, and evolving economic trends could
challenge the model’s consistency and accuracy. Further-
more, its ability to maintain interpretability and resist
biases across varied demographics is crucial for its reliable
deployment in heterogeneous environments.
Neural Computing and Applications
123
6 Conclusion and future directions
This research proposes a novel algorithm for credit card
default prediction by integrating deep learning with XAI
techniques to enhance the interpretability of the decision-
making process. The performance metrics of CreditNetXAI
highlight its superiority, achieving an accuracy of 0.8350,
sensitivity of 0.8823, and specificity of 0.9879. When
juxtaposed against established models like XGB, SVM,
RF, and NN, CreditNetXAI consistently demonstrates
enhanced predictive capabilities across all metrics. This
underscores the model’s competitive edge in prediction
accuracy and its ability to provide profound insights into
credit card default risk determinants. Our work augments
the burgeoning literature on explainable AI within the
financial domain and presents a pragmatic approach to
credit risk assessment, harmonizing accuracy with inter-
pretability. Building upon the foundation of CreditNetXAI,
several avenues for future exploration emerge. The
robustness of CreditNetXAI across varied financial con-
texts and datasets warrants further investigation. Enhanced
feature engineering could capture more nuanced financial
behaviors, refining the model’s predictive capabilities.
Additionally, the algorithm can adapt to address related
challenges in credit risk management, such as predicting
loan defaults or assessing the creditworthiness of new
market entrants. In the future, the proposed algorithm can
be used with OCNN [25–32] and make use of Resnet [33].
ERNN can be used for stress detection as in [34]. Attention
mechanism can be used as in [35] and correlation algo-
rithms as in [36]. YOLO v8 can be used as in [37]. The
proposed algorithm can be adapted to address various
agricultural challenges and play a pivotal role in the sus-
tainable and efficient cultivation of crops worldwide.
Additionally, future work can explore the integration of
CreditNetXAI with state-of-the-art techniques such as
those demonstrated in references [38–41], offering even
more sophisticated capabilities for agricultural decision
support.
Author contributions Fatma, Abdussalam, Mahmoud, and Mostafa
collaborated collaboratively. Fatma and Mostafa came up with the
idea and wrote the abstract and the proposal, while Mahmoud and
Abdussalam contributed by comparing and doing the experiments.
Funding The authors received no specific funding for this study.
Data availability UCI Machine Learning Repository. (2013) [15].
Default of Credit Card Clients Dataset. Retrieved from https://
archive.ics.uci.edu/ml/datasets/default?of?credit?card?clients#.
Declarations
Conflict of interest The authors declare no conflicts of interest to
report regarding the present study.
Ethical approval There are no ethical conflicts.
References
1. Alkhatib K, Al-Aiad A, Almahmood M, Elayan O (2021) Credit
card fraud detection based on deep neural network approach. In:
Proceedings of the 2021 international conference on intelligent
computing and systems (ICICS). IEEE, pp 153–156. https://doi.
org/10.1109/ICICS52457.2021.9464555
2. Baesens B, Roesch D, Scheule H (2016) Credit risk analytics:
measurement techniques, applications, and examples in SAS.
Wiley
3. Chen T, Xu Y, Li X (2021) An interpretable gradient boosting
model for credit card default prediction. J Risk Financ Manag
14(7):319
4. Chen Y, Zhang R (2021) Research on credit card default pre-
diction based on k-means SMOTE and BP neural network.
Complexity 2021:6618841. https://doi.org/10.1155/2021/
6618841
5. Consumer Financial Protection Bureau (n.d.) What happens if I
miss a credit card payment or pay late? https://www.con
sumerfinance.gov/ask-cfpb/what-happens-if-i-miss-a-credit-card-
payment-or-pay-late-en-42/
6. Doshi-Velez F, Kim B (2017) Towards a rigorous science of
interpretable machine learning. https://doi.org/10.48550/arXiv.
1702.08608
7. Federal Reserve Bank of New York (2021) Household debt and
credit report. https://www.newyorkfed.org/microeconomics/hhdc
8. Gao X, Xiong Y, Xiong Z, Xiong H (2022) Credit default risk
prediction based on deep learning. https://doi.org/10.21203/rs.3.
rs-724813/v1
9. Gunning D, Aha D (2019) DARPA’s explainable artificial
intelligence (XAI) program. AI Mag 40(2):53–68. https://doi.org/
10.1609/aimag.v40i2.2850
10. Misheva BH, Osterrieder J, Hirsa A, Kulkarni O, Lin SF (2021)
Explainable AI in credit risk management. https://doi.org/10.
48550/arXiv.2103.00949
11. Husejinovic A, Kec
ˇo D, Mas
ˇetic
´Z (2018) Application of
machine learning algorithms in credit card default payment pre-
diction. Int J Sci Res 7(10):425. https://doi.org/10.15373/
22778179
12. Onay C, O
¨ztu
¨rk E (2018) A review of credit scoring research in
the age of Big Data. J Financ Regul Compliance 26:00–00.
https://doi.org/10.1108/JFRC-06-2017-0054
13. Ribeiro MT, Singh S, Guestrin C (2016) Why should I trust You?
Explaining the predictions of any classifier. Doi: https://doi.org/
10.48550/arXiv.1602.04938
Neural Computing and Applications
123
14. Shi Si, Tse R, Luo W, D’Addona S, Pau G (2022) Machine
learning-driven credit risk: a systemic review. Neural Comput
Appl. https://doi.org/10.1007/s00521-022-07472-2
15. UCI Machine Learning Repository (2013) Default of credit card
clients dataset. https://archive.ics.uci.edu/ml/datasets/
default?of?credit?card?clients#
16. Yeh IC, Lien CH (2009) The comparisons of data mining tech-
niques for the predictive accuracy of probability of default of
credit card clients. Expert Syst Appl 36(2):2473–2480
17. Zhang H, Zheng Y, Chen X (2019) Predicting credit card default
using convolutional neural networks. Expert Syst Appl
125:200–209
18. https://www.clearlypayments.com/blog/cash-use-is-declining-
while-credit-card-use-is-growing/
19. de Lange PE, Melsom B, Vennerød CB, Westgaard S (2022)
Explainable AI for credit assessment in banks. J Risk Financ
Manag 15(12):556. https://doi.org/10.3390/jrfm15120556
20. Ali S, Abuhmed T, El-Sappagh S, Muhammad K, Alonso-Moral
JM, Confalonieri R, Guidotti R, Del Ser J, Dı
´az-Rodrı
´guez N,
Herrera F (2023) Explainable artificial intelligence (XAI): what
we know and what is left to attain trustworthy artificial intelli-
gence. Inf Fusion 99:101805
21. Weber P, Carl KV, Hinz O (2023) Applications of explainable
artificial intelligence in finance—a systematic review of finance,
information systems, and computer science literature. Manag Rev
Q. https://doi.org/10.1007/s11301-023-00320-0
22. Doe J, Smith A, Johnson B (2023) Forecasting the real estate
housing prices using a novel deep learning machine model. Civil
Eng J 9:46–64
23. Smith A, Doe J, Johnson B (2022) Deep learning based convo-
lutional neural network structured new image classification
approach for eye disease identification. Sci Iran. https://doi.org/
10.24200/sci.2022.58049.5537
24. Johnson B, Doe J, Smith A (2023) Chaotic time series recogni-
tion: a deep learning model inspired by complex systems char-
acteristics. Int J Eng 36(1):1–9
25. Talaat FM (2022) Effective deep Q-networks (EDQN) strategy
for resource allocation based on optimized reinforcement learn-
ing algorithm. Multimed Tools Appl. https://doi.org/10.1007/
s11042-022-13000-0
26. Talaat FM (2022) Effective prediction and resource allocation
method (EPRAM) in fog computing environment for smart
healthcare system. Multimed Tools Appl 81(6):8235–8258
27. Talaat Fatma M, Samah A, Nasr Aida A (2022) A new reliable
system for managing virtualcloud network. Comput Mater Con-
tinua 73(3):5863–5885
28. El-Rashidy N, ElSayed NE, El-Ghamry A, Talaat FM (2022)
Prediction of gestational diabetes based on explainable deep
learning and fog computing. Soft Comput 26(21):11435–11450
29. El-Rashidy N, Ebrahim N, el Ghamry A, Talaat FM (2022)
Utilizing fog computing and explainable deep learning tech-
niques for gestational diabetes prediction. Neural Comput Applic.
https://doi.org/10.1007/s00521-022-08007-59
30. Hanaa S, Fatma BT (2022) Detection and classification using
deep learning and sine-cosine fitnessgrey wolf optimization.
Bioengineering 10(1):18. https://doi.org/10.3390/
bioengineering10010018
31. Talaat FM (2023) Real-time facial emotion recognition system
among children with autism based on deep learning and IoT.
Neural Comput Appl. https://doi.org/10.1007/s00521-023-08372-
9
32. Talaat FM (2023) Crop yield prediction algorithm (CYPA) in
precision agriculture based on IoT techniques and climate chan-
ges. Neural Comput Appl. https://doi.org/10.1007/s00521-023-
08619-5
33. E Hassan, N El-Rashidy, FM Talaat (2022) ‘‘Review: Mask
R-CNN Models’’. https://doi.org/10.21608/njccs.2022.280047.
34. Siam AI, Gamel SA, Talaat FM (2023) Automatic stress detec-
tion in car drivers based on non-invasive physiological signals
using machine learning techniques. Neural Comput & Appl.
https://doi.org/10.1007/s00521-023-08428-w
35. Talaat FM, Gamel SA (2023) A2M-LEUK: attention-augmented
algorithm for blood cancer detection in children. Neural Comput
Appl. https://doi.org/10.1007/s00521-023-08678-8
36. Gamel SA, Hassan E, El-Rashidy N et al (2023) Exploring the
effects of pandemics on transportation through correlations and
deep learning techniques. Multimed Tools Appl. https://doi.org/
10.1007/s11042-023-15803-1
37. Talaat FM, ZainEldin H (2023) An improved fire detection
approach based on YOLO-v8 for smart cities. Neural Comput &
Applic. https://doi.org/10.1007/s00521-023-08809-1
38. Alnaggar M, Siam AI, Handosa M, Medhat T, Rashad MZ (2023)
Video-based real-time monitoring for heart rate and respiration
rate. Expert Syst Appl 1(225):120135
39. Alnaggar M, Handosa M, Medhat T, Rashad MZ (2023) Thyroid
disease multi-class classification based on optimized gradient
boosting model. Egypt J Artif Intell 2(1):1–4
40. Alnaggar M, Handosa M, Medhat T, Rashad MZ (2023) An IoT-
based framework for detecting heart conditions using machine
learning. Int J Adv Comput Sci Appl. https://doi.org/10.14569/
IJACSA.2023.0140442
41. Alhussan AA, Talaat FM, El-kenawy ES, Abdelhamid AA,
Ibrahim A, Khafaga DS, Alnaggar M (2023) Facial expression
recognition model depending on optimized support vector
machine. Comput Mater Continua. https://doi.org/10.32604/cmc.
2023.039368
Publisher’s Note Springer Nature remains neutral with regard to
jurisdictional claims in published maps and institutional affiliations.
Springer Nature or its licensor (e.g. a society or other partner) holds
exclusive rights to this article under a publishing agreement with the
author(s) or other rightsholder(s); author self-archiving of the
accepted manuscript version of this article is solely governed by the
terms of such publishing agreement and applicable law.
Neural Computing and Applications
123
Authors and Affiliations
Fatma M. Talaat
1,2,3
•Abdussalam Aljadani
4
•Mahmoud Badawy
5,6
•Mostafa Elhosseini
6,7
&Fatma M. Talaat
fatma.nada@ai.kfs.edu.eg
Abdussalam Aljadani
ajadani@taibahu.edu.sa
Mahmoud Badawy
engbadawy@mans.edu.eg
Mostafa Elhosseini
melhosseini@mans.edu.eg
1
Faculty of Artificial Intelligence, Kafrelsheikh University,
Kafrelsheikh 33516, Egypt
2
Faculty of Computer Science & Engineering, New Mansoura
University, Gamasa 35712, Egypt
3
Nile Higher Institute for Engineering and Technology,
Mansoura, Egypt
4
Department of Management, College of Business
Administration in Yanbu, Taibah University,
41411 Al-Madinah Al-Munawarah, Saudi Arabia
5
Computer Science and Information Department, Applied
College, Taibah University, 46537 Madinah, Saudi Arabia
6
Computers and Control Systems Engineering Department,
Faculty of Engineering, Mansoura University,
Mansoura 35516, Egypt
7
College of Computer Science and Engineering, Taibah
University, 46421 Yanbu, Saudi Arabia
Neural Computing and Applications
123
A preview of this full-text is provided by Springer Nature.
Content available from Neural Computing and Applications
This content is subject to copyright. Terms and conditions apply.
