Journal of Information System Exploration and Research Vol. 2, No. 1, January 2024, pp.

11-20

Journal of Information System

Exploration and Research
https://shmpublisher.com/index.php/joiser
p-ISSN 2964-1160 | e-ISSN 2963-6361

The Optimization of Credit Scoring Model Using Stacking

Ensemble Learning and Oversampling Techniques

Rofik1*, Reza Aulia2, Khalimah Musaadah3, Salma Shafira Fatya Ardyani4, Ade Anggian Hakim5
1Department of Computer Science, Faculty of Mathematics and Natural Science, Universitas Negeri Semarang, Indonesia
2Department of Science, Graduate School of Integrated Science and Technology, Shizouka Univesity, Japan
3Department of Informatics, Faculty of Engineering, Universitas Jenderal Soedirman, Indonesia
4Department of Informatics Engineering, Faculty of Computer Science, Universitas Dian Nuswantoro, Indonesia
5Department of Electrical Engineering, Faculty of Electrical and Electronic Engineering, University Tun Hussein Onn Malaysia,

Malaysia

DOI: https://doi.org/10.52465/joiser.v2i1.203
Received 27 August 2023; Accepted 30 October 2023; Available online xx December 2023

Article Info Abstract

Keywords: Credit risk assessment plays an important role in efficient and safe
Credit scoring; banking decision-making. Many studies have been conducted to analyze
Stacking ensemble learning;
credit scoring with a focus on achieving high accuracy. However,
XGBoost;
predicting credit scoring decisions also requires model construction that
SMOTE
handles class imbalance and proper model implementation. This
research aims to increase the accuracy of credit assessment by balancing
data using Synthetic Minority Oversampling (SMOTE) and applying
ensemble stacking learning techniques. The proposed model utilizes a
base learner consisting of Random Forest, SVM, Extra-Tree Classifier, and
XGboost as a meta-learner. Then to handle unbalanced classes using
SMOTE. The research process was carried out in several stages, namely
Data Collection, Preprocessing, Oversampling, Modeling, and Evaluation.
The model was tested using the German Credit dataset by applying cross-
validation. The evaluation results show that the stacking ensemble
learning model developed has optimal performance, with an accuracy of
83.21%, precision of 79.29%, recall of 91.78%, and f1-score of 85.08%.
This research shows that optimizing the stacking ensemble learning
model with data balancing using SMOTE in credit scoring can improve
performance in credit scoring.

This is an open-access article under the CC BY-SA license.

1. Introduction

In the current era of information technology development and financial innovation, the credit
scoring process has become a crucial cornerstone in the decision-making of banks and lending

*
Corresponding Author:
Rofik,
Department of Computer Science,
Faculty of Mathematics and Natural Science, Universitas Negeri Semarang,
Semarang, Indonesia.
Email: [email protected]

11
institutions [1]. Credit scoring is a job that aims to assess the credit risk of a prospective customer or
business entity, which helps financial institutions make efficient and accurate decisions regarding the
granting of credit [2]. In applying data analysis techniques and predictive models, credit scoring plays
an important role in maintaining the balance between safe financing and lending to the right parties.
This is because inappropriate lending decisions can result in huge losses [2].
Lending behavior and technological developments have changed the landscape of the credit
assessment process [3]. Especially in this digital era, online lending has become a popular alternative
for individuals and businesses to obtain funds without involving traditional financial institutions such
as banks [4]. Such loans utilize the convenience and transparency of online platforms, resulting in a
faster and easier process compared to conventional approaches [5]. However, online consumer lending
is of great concern to the Company, as online consumer lending typically carries a higher credit risk than
businesses in the conventional lending system [6].
Based on information from the Wind economic database in 2019, in China, the accumulated rate of
1-year payment defaults of some lending institutions or securities backing online consumer loans even
reached 9% [7]. Although it is difficult to ensure that a borrower will default in the future, with advanced
analytical approaches and appropriate modeling techniques, credit scoring can indicate potential
default risk before a credit transaction is made [8], [9].
However, this development also brings about new challenges in credit risk assessment. In addition
to the disadvantages of the traditional approach that can be easily affected by sample selection bias,
as it only uses a sample of accepted applicants, while the applicant population also includes rejected
applicants [10], [11]. The existence of large and highly variable data, as well as the complexity of factors
that can affect credit decisions, makes demands on lending institutions to still be able to perform credit
scoring quickly and accurately [12]. Therefore, many researchers have developed techniques to assess
credit risk with data mining [11], [13]–[16]. Data mining is a machine learning method that aims to
extract valuable information from existing data [3]. Using this technique, important information can be
identified from customer data, credit history, and other relevant factors.
There have been many studies that implement data mining in credit scoring. Research by [17] who
did a new approach to assessing credit applications by giving a binary score, by combining a Genetic
Algorithm (GA) with Support Vector Machine (SVM). By applying 2 levels, which are determining the
SVM parametrization and finding the most weighted feature set, this research was able to achieve an
accuracy of 80.70%. Olivares et al in their research [18] explored the application of discrete-time joint
models in credit scoring. The study combined survival analysis with longitudinal data by integrating
variable covariates in survival analysis. From the study, it was found that the inclusion of time-varying
covariates in the survival model improved the prediction of credit scoring. Using Australian, German,
and Japanese datasets, the research in [19] focuses on the implementation of the Extreme Learning
Machine (ELM) classification tool for the credit scoring analysis model. Since ELM requires more hidden
neurons and random determination of input weights and hidden biases, the study proposed a novel
activation function and evolutionary approach to obtain optimized weights and biases using the Bat
algorithm. From the model that has been built, this research can achieve consecutive accuracy of the
Australian, German, and Japanese datasets of 89.92%, 81.18%, and 88.35%. The research conducted by
[12] is focused on a new development called soft reordering one-dimensional CNN (SR-1D-CNN) which
is designed to adaptively restructure the original table data to better suit CNN learning. By using 5
datasets from Polish, Ashare, GiveMeSomeCredit, Lending Club and HomerCreditDefaultRisk, the
model built was able to produce the greatest accuracy of 95.18 from the Polish dataset.
The research conducted by the author will focus on developing a credit-scoring model using stacking
ensemble learning techniques and handling unbalanced data. Despite technological advances and the
application of machine learning models in credit scoring analysis, the main problem that often arises is
the inability of the model to explain predictions and data imbalance [20]. As done in [9], [21]–[23] which
focuses on handling class imbalance. It can be concluded that the performance of the model decreases
inversely with increasing the level of class imbalance. This is proven by [1] by evaluating class imbalance
using LIME and SHAP stability. This shows that the resulting interpretations of LIME and SHAP are less
stable as class imbalance increases, concluding that class imbalance does hurt machine learning
interpretations. Research conducted by [24]–[26] also revealed that ensemble models tend to produce
good performance. Research [27] also reveals that in the context of credit scoring, ensemble methods
based on decision trees such as random forest algorithms produce better classification performance
compared to standard logistic regression models.

12
2. Method

In this research, credit scoring analysis is carried out with several stages, namely, data collection,
preprocessing, oversampling, modeling, and evaluation. The framework of the stages of this research
can be seen in Figure 1.

Figure 1. Research Framework of Credit Scoring

A comprehensive framework is proposed as a guide to understanding the process that must be done
to achieve high-accuracy credit scoring prediction results. The framework incorporates various
important stages starting from data collection to evaluation. Figure 1 shows the main steps taken in
this research. Through these steps, it is expected to present a solution that can overcome the problems
that usually arise in credit scoring including data imbalance. A more detailed explanation of each stage
in this research framework can be seen as follows.

2.1 Data Collection

In this section, data related to customer credit is grouped. The dataset used in this research is a
German dataset sourced from UCI Machine Learning. The dataset consists of 20 attributes, i.e. "status",
"duration", "credit_history", "purpose", "amount", "savings", "tenure", "installment_amount",
"installment_level", "status_gender", "debtor_other", "current_place_of_residence", "property",
"age", "installment plan_other", "place_of_residence", "amount_of_credit", "occupation",
"amount_of_responsibility", "telephone", "foreign_worker", and "credit_risk" as the labels. The
dataset consists of 1000 records of credit data. Of these data, 700 records are good classes and 300 are
bad classes. The dataset used in this research can be accessed via the URL link:
https://archive.ics.uci.edu/dataset/573/south+german+credit+update.

2.2 Data preprocessing

Data preprocessing is used to clean and prepare the dataset so that it can be implemented and
support the modeling stage. Data separation is done to separate features (X) which are features in the
credit scoring dataset, and target variables (Y) which are class data. Feature encoding is also done using
One-Hot Encoding. Data standardization is also done using Standard Scaler. This is done to make all
data in each feature have the same scale and avoid model sensitivity issues due to scale differences.

2.3 Data oversampling

Oversampling of data is done to overcome datasets that have unbalanced classes [28]. Where the
number of good consumers is more than the number of bad consumers. In this case are accepted credit
applicants and rejected credit applicants, to overcome this problem which will have a habitual impact
on the model, the oversampling technique is carried out [29]. With the application of oversampling, the
model built will not be more inclined to the majority class only during the training process but is good
at generalizing both classes. The oversampling technique performed in this research is the Synthetic

13
Minority Over-sampling Technique Evaluation (SMOTE) method. This technique is capable of generating
synthesized data [30]. The sample is generated by increasing the sample that is different from the
minority class sample [31]. SMOTE is a data augmentation method technique for classification datasets,
which improves recognition performance without increasing the risk of data leakage [32].

2.4 Modeling with Stacking Ensemble Learning.

Initialization of the base model is done by selecting the base model used. In this research, Random
Forest, SVM, and Extra-Tree Classifier algorithms were used. Some of the reasons for choosing these
algorithms as base models include Random Forest, with its ability to overcome overfitting [33] and can
produce stable and accurate predictions, which is used to combine several decision trees to produce
the final prediction. Where the prediction results are drawn through voting from decision trees that
work independently. It also outperforms logistic regression algorithms and is developing into a major
algorithm in the credit scoring sector [34]. The SVM algorithm is also chosen and applied as a base
model, due to its popularity and efficiency in solving classification and regression problems [35]. The
working concept of this algorithm is to separate two classes by maximizing the margin between the two
classes. Meanwhile, the Extra-Tree Classifier is also applied as a base model because in some studies it
has been proven to show good performance [36]. This algorithm belongs to the ensemble learning
category which is similar to Random Forest. However, this algorithm works by building a decision tree
with random features and subsampling. Its advantage is that it is suitable for large data and its
computational speed is also high.
These algorithms are trained using training data and can understand patterns in the data and make
predictions. Once trained, the algorithms are used to predict test data, where each model predicts the
probability for each class. The probability results from each base model are used by the final estimator
or in this study, XGBoost. The ensemble uses the probability results from the base models as input
features to help the meta-model produce better final predictions.
Previously the data was divided into training data and testing data. The ratio is 80% training data and
20% testing data. A cross-validation of 5 times was performed on the ensemble stacking model
developed using the 80% training data. This method is done because it has a good impact on the model
built [37]. The concept is to train the model by using alternating data as training and testing, which from
this process can help the model generalize data that is not seen in training. The prediction results
provide a probability value for each credit risk class which is then used to determine the final decision.
From these probability values, various thresholds that separate high and low credit risk classes can be
explored. This research also tested several thresholds from the range of 0.1 to 0.9 in cross-validation.
Threshold testing is done in a separate loop after all cross-validation iterations are completed. The
threshold that yields the greatest accuracy among other threshold tests is used to classify the prediction
results.

2.5 Evaluation

The confusion metric is one of the evaluation metrics used to analyze the performance of the stacking
ensemble learning model, which is also used in this study. The confusion matrix can provide an overview
of the prediction and actual state given by the algorithm model. Confusion Matrix has 4 important
elements. Among them is True Positives (TP), which represents the amount of data that is actually in
the positive class and predicted by the model as a positive class. Then True Negative (TN), which reflects
the amount of data that is actually in the negative class and is predicted by the model as a negative
class. While False Positive (FP) is the amount of data that is actually in the negative class but is predicted
by the model as a positive class. False Negative (FN) is the amount of data that is actually in the positive
class but predicted by the model as a negative class. By using the Confusion matrix, we can calculate
and evaluate the performance of the model that has been built through metrics such as accuracy,
precision, recall, and f1-score. The calculation details can be seen below.

1. Accuracy
Accuracy is a value that indicates how accurate the model is in predicting the entire data. Measured
by the formula:
(𝑇𝑃 + 𝑇𝑁) (1)
𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 =
(𝑇𝑃 + 𝑇𝑁 + 𝐹𝑃 + 𝐹𝑁)

14
2. Precision
Precision measures the degree to which work predicted to be fake is fake. We can calculate it with the
following formula:
𝑇𝑃 (2)
𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 =
(𝑇𝑃 + 𝐹𝑃)

3. Recall (Sensitivity)
Recall measures the extent to which the model successfully detects fake jobs overall. Recall calculated
by the following formula:
𝑇𝑃 (3)
𝑅𝑒𝑐𝑎𝑙𝑙 =
(𝑇𝑃 + 𝐹𝑁)

4. F1-Score
The F1-Score is a combination of precision and recall into a single metric that yields the overall model
performance. This can be calculated by the following formula:
2 ∗ (𝑃𝑟𝑒𝑠𝑖𝑠𝑖 ∗ 𝑅𝑒𝑐𝑎𝑙𝑙) (4)
𝐹1 − 𝑆𝑐𝑜𝑟𝑒 =
(𝑃𝑟𝑒𝑠𝑖𝑠𝑖 + 𝑅𝑒𝑐𝑎𝑙𝑙)

3. Results and Discussion

This research uses the German credit dataset that has been used in previous studies. The research
was carried out in stages starting with data collection, namely collecting datasets where datasets used
were obtained from UCI Machine Learning. Perform preprocessing, separating features (X) and target
variables (Y). Perform data coding using One-Hot Encoding, and standardize data using StandardScaler.
Then oversampling the data using SMOTE to overcome data imbalances that can cause the model's
performance to be not optimal in classification. The modeling stage is carried out to build a stacking
ensemble learning model that combines Random Forest, SVM, and Extra-Tree Classifier algorithms as
base learner. And Xgboost as a meta-learner model. The next stage is evaluation, cross-validation is
done by calculating the probability of prediction using the stacking model. Apply to determine the Best
Threshold to maximize accuracy by iterating through several threshold values. Whichever threshold is
best is used to classify the data and calculate cross-validation accuracy. Then the model performance
is calculated through accuracy, precision, recall, and f1-score. The following are the threshold
experiments performed on cross-validation that correlate with the accuracy obtained.

The results of oversampling to balance the data using the SMOTE method against the amount of
data can be seen in Figure 2.

(a) (b)

Fig 2. Data distribution (a) before oversampling and (b) after oversampling.

Handling unbalanced classes is done using the SMOTE method which augments the data from the
minority class, resulting in data that is balanced between the classes. This reduces bias in predictions
and allows the model to learn better patterns from both classes in the dataset. As a result of the SMOTE

15
implementation, there is a class balance of 700, whereas previously the bad credit risk class was 300
and the good credit risk class was 700.
A heatmap is displayed that shows the correlation between the features in the dataset and its
target, which in this case is 'credit_risk'. The darker the color, the stronger the correlation, and vice
versa. The value that shows the size of the correlation if it is close to 1 means a positive correlation,
while otherwise it is negatively correlated. Whereas when the value is 0, it means that there is no
correlation between the two features.

Figure 3. Heatmap of correlation between features

The 3 features that are strongly positively correlated with the target include status (checking account
status), credit history, and savings. The most negatively correlated features include duration (length of
loan), amount (amount of money borrowed), and property (cars, real property, buildings, and so on).
The experimental analysis of several thresholds from 0.1 to 0.9 on the credit score prediction model
using the stacking ensemble learning technique can be seen in Figure 4.

16
 Figure 4. Model performance on testing each threshold

In each cross-validation fold, the model is fit to the training which is then used to predict the
probability of a positive class. Different thresholds are also applied. From the figure, it can be seen that
the average accuracy changes as the threshold moves from 0.1 to 0.9. Each point on the grid represents
the average accuracy of the model across all cross-validation folds at a particular threshold. The largest
average accuracy obtained was 83.21% with a threshold of 0.2.
This research uses a stacking ensemble learning model by combining 3 algorithms, namely Random
Forest, SVM, and Extra-Tree Classifier. The results obtained state that the model built successfully
produces good performance and can improve the accuracy performance of previous research. The test
results were evaluated using a confusion matrix in the form of accuracy, precision, recall, and score.
The model built was able to produce the greatest accuracy of 83.21%, precision of 79.29%, recall of
91.78%, and F1-score of 85.07%. A comparison of the performance of the ensemble stacking model
results built in this study with previous research models can be seen in Table 1.

Table 1. Comparison of techniques and results from previous research

Approach Techniques or algorithms used Year Accuracy
[38] Combines the benefits of feature selection and 2018 77.12%
ensemble frameworks. Using 5 basic
classification algorithms and combining them
with the weighted voting approach
[39] Bayesian optimization and PSO 2019 78.30%
implementation
[17] Combination of genetic algorithm and SVM 2020 80.70%
[40] Multi-grained augmented gradient boosting 2021 77.15%
decision trees (mg-GBDT), (Gradient boosting
decision trees, Multi-grained scanning)
Proposed Method Optimization of Credit Scoring Analysis Model 2023 83.21%
using Stacking Ensemble Learning Approach
and Oversampling with SMOTE

4. Conclusion

In this research, credit scoring classification between good credit risk and bad credit risk is carried
out using stacking ensemble learning from Random Forest, SVM, and Extra-Tree Classifier algorithms.

17
The meta-learner model is XGBoost. This research shows the effectiveness of the stacking ensemble
learning model in classifying good and bad credit risk, namely between accepted and unaccepted credit
applicants. Oversampling using SMOTE is used to overcome unbalanced datasets. An inter-class
boundary search was also conducted using thresholds from 0.1 to 0.9 for classification. The evaluation
results showed that the ensemble stacking model successfully improved the performance in
distinguishing good and bad credit risk. The resulting performance of the stacking model achieved the
best accuracy of 83.21% with a precision of 79.28%, recall of 91.78%, and f1-score of 85.97%. And the
best threshold to separate the two classes is 0.2. In future research, it is recommended to further
explore larger data, perform feature selection, and try to create new models to achieve more optimal
credit scoring performance.

References

[1] Y. Chen, R. Calabrese, and B. Martin-Barragan, “Interpretable machine learning for imbalanced
credit scoring datasets,” Eur. J. Oper. Res., no. xxxx, 2023, doi: 10.1016/j.ejor.2023.06.036.
[2] Z. Zhang, Y. Li, Y. Liu, and S. Liu, “A local binary social spider algorithm for feature selection in
credit scoring model,” Appl. Soft Comput., vol. 144, p. 110549, 2023, doi:
10.1016/j.asoc.2023.110549.
[3] M. A. Muslim et al., “New model combination meta-learner to improve accuracy prediction P2P
lending with stacking ensemble learning,” Intell. Syst. with Appl., vol. 18, no. December 2022,
p. 200204, 2023, doi: 10.1016/j.iswa.2023.200204.
[4] X. Ma, J. Sha, D. Wang, Y. Yu, Q. Yang, and X. Niu, “Study on a prediction of P2P network loan
default based on the machine learning LightGBM and XGboost algorithms according to different
high dimensional data cleaning,” Electron. Commer. Res. Appl., vol. 31, no. February, pp. 24–
39, 2018, doi: 10.1016/j.elerap.2018.08.002.
[5] X. Fu, S. Zhang, J. Chen, T. Ouyang, and J. Wu, “A Sentiment-Aware Trading Volume Prediction
Model for P2P Market Using LSTM,” IEEE Access, vol. 7, pp. 81934–81944, 2019, doi:
10.1109/ACCESS.2019.2923637.
[6] M. Di Maggio and V. Yao, “Fintech Borrowers: Lax Screening or Cream-Skimming?,” Rev. Financ.
Stud., vol. 34, no. 10, pp. 4565–4618, 2021, doi: 10.1093/rfs/hhaa142.
[7] Y. Xia, Y. Li, L. He, Y. Xu, and Y. Meng, “Incorporating multilevel macroeconomic variables into
credit scoring for online consumer lending,” Electron. Commer. Res. Appl., vol. 49, no. March,
p. 101095, 2021, doi: 10.1016/j.elerap.2021.101095.
[8] C. Bai, B. Shi, F. Liu, and J. Sarkis, “Banking credit worthiness: Evaluating the complex
relationships,” Omega (United Kingdom), vol. 83, pp. 26–38, 2019, doi:
10.1016/j.omega.2018.02.001.
[9] F. Shen, X. Zhao, G. Kou, and F. E. Alsaadi, “A new deep learning ensemble credit risk evaluation
model with an improved synthetic minority oversampling technique,” Appl. Soft Comput., vol.
98, p. 106852, 2021, doi: 10.1016/j.asoc.2020.106852.
[10] Y. Kang, N. Jia, R. Cui, and J. Deng, “A graph-based semi-supervised reject inference framework
considering imbalanced data distribution for consumer credit scoring,” Appl. Soft Comput., vol.
105, p. 107259, 2021, doi: 10.1016/j.asoc.2021.107259.
[11] R. A. Mancisidor, M. Kampffmeyer, K. Aas, and R. Jenssen, “Deep generative models for reject
inference in credit scoring,” Knowledge-Based Syst., vol. 196, p. 105758, 2020, doi:
10.1016/j.knosys.2020.105758.
[12] H. Qian, P. Ma, S. Gao, and Y. Song, “Soft reordering one-dimensional convolutional neural
network for credit scoring,” Knowledge-Based Syst., vol. 266, p. 110414, 2023, doi:
10.1016/j.knosys.2023.110414.
[13] Y. Wang, Y. Jia, Y. Zhong, J. Huang, and J. Xiao, “Balanced incremental deep reinforcement
learning based on variational autoencoder data augmentation for customer credit scoring,”
Eng. Appl. Artif. Intell., vol. 122, no. March, p. 106056, 2023, doi:
10.1016/j.engappai.2023.106056.
[14] H. He, Z. Wang, H. Jain, C. Jiang, and S. Yang, “A privacy-preserving decentralized credit scoring
method based on multi-party information,” Decis. Support Syst., vol. 166, no. November 2022,
p. 113910, 2023, doi: 10.1016/j.dss.2022.113910.
[15] Y. Wu, W. Huang, Y. Tian, Q. Zhu, and L. Yu, “An uncertainty-oriented cost-sensitive credit

18
 scoring framework with multi-objective feature selection,” Electron. Commer. Res. Appl., vol.
53, no. March, p. 101155, 2022, doi: 10.1016/j.elerap.2022.101155.
[16] D. M. B. Silva, G. H. A. Pereira, and T. M. Magalhães, “A class of categorization methods for
credit scoring models,” Eur. J. Oper. Res., vol. 296, no. 1, pp. 323–331, 2022, doi:
10.1016/j.ejor.2021.04.029.
[17] D. Şen, C. Ç. Dönmez, and U. M. Yıldırım, “A Hybrid Bi-level Metaheuristic for Credit Scoring,”
Inf. Syst. Front., vol. 22, no. 5, pp. 1009–1019, 2020, doi: 10.1007/s10796-020-10037-0.
[18] V. Medina-Olivares, R. Calabrese, J. Crook, and F. Lindgren, “Joint models for longitudinal and
discrete survival data in credit scoring,” Eur. J. Oper. Res., vol. 307, no. 3, pp. 1457–1473, 2023,
doi: 10.1016/j.ejor.2022.10.022.
[19] D. Tripathi, D. R. Edla, V. Kuppili, and A. Bablani, “Evolutionary Extreme Learning Machine with
novel activation function for credit scoring,” Eng. Appl. Artif. Intell., vol. 96, no. February, p.
103980, 2020, doi: 10.1016/j.engappai.2020.103980.
[20] X. Dastile, T. Celik, and M. Potsane, “Statistical and machine learning models in credit scoring:
A systematic literature survey,” Appl. Soft Comput. J., vol. 91, p. 106263, 2020, doi:
10.1016/j.asoc.2020.106263.
[21] Y. Li, T. Bellotti, and N. Adams, “Issues Using Logistic Regression With Class Imbalance, With a
Case Study From Credit Risk Modelling,” Found. Data Sci., vol. 1, no. 4, pp. 389–417, 2019, doi:
10.3934/fods.2019016.
[22] K. Zhang et al., “Label correlation guided borderline oversampling for imbalanced multi-label
data learning,” Knowledge-Based Syst., vol. 279, p. 110938, 2023, doi:
10.1016/j.knosys.2023.110938.
[23] X. Tao, X. Guo, Y. Zheng, X. Zhang, and Z. Chen, “Self-adaptive oversampling method based on
the complexity of minority data in imbalanced datasets classification,” Knowledge-Based Syst.,
vol. 277, p. 110795, 2023, doi: 10.1016/j.knosys.2023.110795.
[24] G. E-mail, K. W. De Bock, K. Coussement, S. Lessmann, and K. W. De Bock, “Version of Record:
https://www.sciencedirect.com/science/article/pii/S0377221720300898,” pp. 1–39.
[25] W. Yin, B. Kirkulak-Uludag, D. Zhu, and Z. Zhou, “Stacking ensemble method for personal credit
risk assessment in Peer-to-Peer lending,” Appl. Soft Comput., vol. 142, p. 110302, 2023, doi:
10.1016/j.asoc.2023.110302.
[26] P. Sulikowski and T. Zdziebko, “Churn factors identification from real-world data in the
telecommunications industry: Case study,” Procedia Comput. Sci., vol. 192, pp. 4800–4809,
2021, doi: 10.1016/j.procs.2021.09.258.
[27] E. Dumitrescu, S. Hué, C. Hurlin, and S. Tokpavi, “Machine learning for credit scoring: Improving
logistic regression with non-linear decision-tree effects,” Eur. J. Oper. Res., vol. 297, no. 3, pp.
1178–1192, 2022, doi: 10.1016/j.ejor.2021.06.053.
[28] W. F. Abror and M. Aziz, “Journal of Information System Bankruptcy Prediction Using Genetic
Algorithm-Support Vector Machine ( GA-SVM ) Feature Selection and Stacking,” vol. 1, no. 2,
pp. 103–108, 2023.
[29] J. Mushava and M. Murray, “A novel XGBoost extension for credit scoring class-imbalanced data
combining a generalized extreme value link and a modified focal loss function,” Expert Syst.
Appl., vol. 202, no. March, p. 117233, 2022, doi: 10.1016/j.eswa.2022.117233.
[30] A. Fernández, S. García, F. Herrera, and N. V. Chawla, “SMOTE for Learning from Imbalanced
Data: Progress and Challenges, Marking the 15-year Anniversary,” J. Artif. Intell. Res., vol. 61,
pp. 863–905, 2018, doi: 10.1613/jair.1.11192.
[31] J. Sun, J. Lang, H. Fujita, and H. Li, “Imbalanced enterprise credit evaluation with DTE-SBD:
Decision tree ensemble based on SMOTE and bagging with differentiated sampling rates,” Inf.
Sci. (Ny)., vol. 425, pp. 76–91, 2018, doi: 10.1016/j.ins.2017.10.017.
[32] A. Imakura, M. Kihira, Y. Okada, and T. Sakurai, “Another use of SMOTE for interpretable data
collaboration analysis,” Expert Syst. Appl., vol. 228, no. January, 2023, doi:
10.1016/j.eswa.2023.120385.
[33] A. Hoarau, A. Martin, J. C. Dubois, and Y. Le Gall, “Evidential Random Forests,” Expert Syst. Appl.,
vol. 230, no. February, p. 120652, 2023, doi: 10.1016/j.eswa.2023.120652.
[34] Y. Liu, M. Yang, Y. Wang, Y. Li, and T. Xiong, “Applying machine learning algorithms to predict
default probability in the online credit market: Evidence from China,” Int. Rev. Financ. Anal.,
vol. 79, no. November 2021, p. 101971, 2022, doi: 10.1016/j.irfa.2021.101971.
[35] M. A. Ganaie, A. Kumari, A. Girard, J. Kasa-Vubu, and M. Tanveer, “Diagnosis of Alzheimer’s

19
 disease via Intuitionistic fuzzy least squares twin SVM,” Appl. Soft Comput., vol. 149, no.
September, 2023, doi: 10.1016/j.asoc.2023.110899.
[36] H. F. Chen et al., “Predicting residual stress of aluminum nitride thin-film by incorporating
manifold learning and tree-based ensemble classifier,” Mater. Chem. Phys., vol. 295, no. 300,
p. 127070, 2023, doi: 10.1016/j.matchemphys.2022.127070.
[37] S. M. Malakouti, “Improving the prediction of wind speed and power production of SCADA
system with ensemble method and 10-fold cross-validation,” Case Stud. Chem. Environ. Eng.,
vol. 8, no. April, p. 100351, 2023, doi: 10.1016/j.cscee.2023.100351.
[38] D. Tripathi, D. R. Edla, V. Kuppili, A. Bablani, and R. Dharavath, “Credit Scoring Model based on
Weighted Voting and Cluster based Feature Selection,” Procedia Comput. Sci., vol. 132, no.
Iccids, pp. 22–31, 2018, doi: 10.1016/j.procs.2018.05.055.
[39] S. Guo, H. He, and X. Huang, “A Multi-Stage Self-Adaptive Classifier Ensemble Model With
Application in Credit Scoring,” IEEE Access, vol. 7, pp. 78549–78559, 2019, doi:
10.1109/ACCESS.2019.2922676.
[40] W. Liu, H. Fan, and M. Xia, “Step-wise multi-grained augmented gradient boosting decision
trees for credit scoring,” Eng. Appl. Artif. Intell., vol. 97, no. May 2020, p. 104036, 2021, doi:
10.1016/j.engappai.2020.104036.

20