Hiệu quả của các phương pháp xử lý dữ liệu mất cân bằng trong chấm điểm tín dụng: Trường hợp tại các Ngân hàng thương mại Việt Nam

Bài viết này nghiên cứu hiệu quả của phương pháp xử lý dữ liệu mất cân bằng trong bài toán phân loại khách hàng tại các ngân hàng thương mại. Đây là một vấn đề phổ biến trong vấn đề phân loại khách hàng, trong đó các quan sát của một lớp nhiều hơn lớp còn lại trong dữ liệu. Mời các bạn cùng tham khảo!

THE EFFECTIVENESS OF METHODS IN DEALING

WITH IMBALANCED DATA IN CREDIT SCORING: THE CASE

OF VIETNAM COMMERCIAL BANKS HIỆU QUẢ CỦA CÁC PHƯƠNG PHÁP XỬ LÝ DỮ LIỆU

MẤT CÂN BẰNG TRONG CHẤM ĐIỂM TÍN DỤNG: TRƯỜNG HỢP

TẠI CÁC NGÂN HÀNG THƯƠNG MẠI VIỆT NAM

Nguyen Thi Lien, MS ; Nguyen Thi Thu Trang, MS; Nguyen Thi Dung

National Economics University

lientkt@neu.edu.vn

Abstract

This article investigates the effectiveness of imbalanced data processing methods in the problem of customer classification at commercial banks This is a common issue in a customer classification attempt, where observations of one class outnumber the remaining class We apply the methods widely used in the world including undersampling, oversampling, bothsamling tech-niques, and SMOTE (Synthetic Minority Oversampling Technique) to deal with imbalances The logit model is applied to datasets that have been processed by these methods to classify customers Using 7501 transaction data from individual customers, the classification results using data processed with these techniques all improve significantly compared to using untreated data Be-sides, the results also show that the most efficient method is SMOTE technique combined with the logit model using variables transformed by Weight of Evidence (WOE).

Keywords: Bothsampling, credit scoring, oversampling, SMOTE, undersampling, WOE.

Tóm tắt

Bài báo này nghiên cứu hiệu quả của phương pháp xử lý dữ liệu mất cân bằng trong bài toán phân loại khách hàng tại các ngân hàng thương mại Đây là một vấn đề phổ biến trong vấn

đề phân loại khách hàng, trong đó các quan sát của một lớp nhiều hơn lớp còn lại trong dữ liệu Chúng tôi áp dụng các phương pháp được sử dụng rộng rãi trên thế giới bao gồm kỹ thuật lấy mẫu dưới (Undersampling), lấy mẫu quá mức (Oversampling), kỹ thuật lấy mẫu cả hai (booth-sampling) và SMOTE (Synthetic Minority Oversampling Technique) để giải quyết vấn đề mất cân bằng Mô hình logit được áp dụng cho các tập dữ liệu đã được xử lý mất cân bằng để phân loại khách hàng Sử dụng 7501 dữ liệu giao dịch từ các khách hàng cá nhân, kết quả phân loại

sử dụng dữ liệu đã được xử lý mất cân bằng đều cải thiện đáng kể so với sử dụng dữ liệu không được xử lý Bên cạnh đó, kết quả cũng cho thấy phương pháp hiệu quả nhất là kỹ thuật SMOTE kết hợp với mô hình logit sử dụng các biến đầu vào được chuyển đổi sang trọng số bằng chứng (Weight of Evidence - WOE)

Từ khóa: Chấm điểm tín dụng, lấy mẫu cả hai, lấy mẫu quá mức, lấy mẫu dưới, SMOTE,

WOE

1 Introduction

In the case of classification algorithms, we aim to predict the false observations with the highest accuracy However, the accuracy of these algorithms can be influenced by the imbalanced data (López, 2013), where observations of one class outnumber the other class(es) at least 5:1 for the binary case (He & Garcia, 2009) In order to improve the prediction accuracy and lower the computation expense, a previous identification and quantification of the most relevant input variables of the model is always highly advised Imbalanced data also appears in some other fields, such as fault detection (Gong & Qiao, 2012; Silva et al., 2006; Wei et al., 2013), toxin de-tection (Harley et al., 2020), medical diagnosis (Ertekin et al., 2007) and customer churn predic-tion (Bing Zhu et al., 2017) In medical area, consider the group of cancer patients as the positive class and remaining persons are in the negative class, the difference about quantity between these two groups in datasets is large and unequal in the number of observations between the groups Other examples of rare events include software defects (Rodriguez et al., 2014), cancer gene ex-pressions (Wu et al., 2012), credit card transactions fraudulent (Kundu et al., 2009), fraud detec-tion in telecommunicadetec-tion (Augustin, 2012), and natural disaster events (Kim et al., 2016) The most prevalent class is called the majority class, while the rarest class is called the minority class (Huang et al., 2016)

Several techniques to process imbalanced data are developed and applied on the aspect of data while others concentrate on the algorithmic level To address the imbalanced problem, four categories in the applications include the preprocessing strategies, cost-sensitive learning meth-ods, adaptation of machine learning techniques and combination of the previous three approaches The preprocessing strategies include resampling techniques and/or variable importance analysis Resampling algorithms are presented by oversampling and undersampling methods Resampling techniques are used to rebalance the sample space for an imbalanced dataset to alleviate the effect

of the skewed class distribution in the learning process Resampling methods are more versatile because they are independent of the selected classifier (López et al., 2013) Oversampling meth-ods are created the synthetic minority samples are randomly duplicating the minority samples Some research (Chawla et al., 2002; Estabrooks et al., 2004; and Tahir et al., 2009) introduced the random undersampling in managing the imbalanced dataset The undersampling method drops non-default observations to oversee the data imbalance However, those methods have limitations such as overfitting or loosing essential information The SMOTE (Synthetic Minority Oversam-pling Technique) method increases instances for minority class based on KNN (K Nearest Neigh-bor) algorithm introduced by Batista et al (2004) The added observations have data properties that are close to the original observations, reducing the level of imbalance in the data set Based

on the original SMOTE method, several newly developed algorithms include SMOTE-N (Syn-thetic Minority Oversampling Technique Nominal), SMOTE-NC (Syn(Syn-thetic Minority Oversam-pling Technique Nominal Continuous), MWMOTE (Majority Weighted Minority OversamOversam-pling Technique) Sukarna Barua et al (2012) introduced two methods, Borderline-SMOTE 1 and Bor-derline-SMOTE 2, which only alter the observations at the border of the sample to reduce the misclassification rate The SMOTEBoost method (Chawla et al., 2003) improved the SMOTE method by incorporating the AdaBoost M2 algorithm The EasyEnsemble method (proposed by Liu et al., 2009) has increased the efficiency of the undersampling algorithm by adding useful

information to the clustering method to reduce the noise level of the generated sample Besides, the BalanceCascade method (Han et al., 2009) provides a method to guide the deletion of obser-vations rather than the random deletion method Besides, bothsampling is a combination of the above methods (Dubey et al., 2014)

The traditional classification method for predicting the probability of default of a loan is logistic regression (Hosmer et al., 2000; Maalouf and Trafalis, 2011; Shu et al., 2014; Maratea et al., 2014); neural networks and decision trees (Quinlan, 1998; Sarkar et al., 2016) Other methods are applied including gradient boosting, least square support vector machines (Jin et al., 2012) These research models focus on statistical orientation or artificial intelligence (AI)

Input variable analysis of these algorithms has gained attention in many practical applica-tions (Ferretti et al., 2016) due to the complexity of interacapplica-tions among variables on large datasets Variable analysis is a crucial task to improve the model interpretability, reduce the computational cost, optimize the data storage, and provide a smaller number of relevant input variables Several approaches have been considered, which can be grouped input variables into new categories or values without losing explanatory prediction power Principal Component Analysis (PCA) is only suitable for data that are normally distributed, near-standard, or linearly related features (Jolliffe and Cadima, 2016) Other analysis constructs a model base on the Weight of Evidence (WOE) and Information Value (IV) in clustering algorithms (Polykretis and Chalkias, 2018)

In Vietnam, the data processing algorithm by the undersampling method of reducing ran-dom elements on the data boundary is introduced in medical data (Phuong et al., 2015) Research

on SMOTE techniques (Lien et al., 2018) in machine learning techniques in credit card fraud de-tection show that the algorithm is suitable for controlling credit card fraud dede-tection in Vietnam Currently, commercial banks in Vietnam have many difficulties in building internal risk management models according to Basel II capital standards (issued by the Basel Committee on Banking Supervision (BIS,2004)) and Circular No.41/2016/TT-NHNN (issued by the State Bank

of Vietnam) Specific difficulties occur in the process of data collection, data processing, selection and evaluation of the effectiveness of the internal risk management model More seriously, ac-cording to VietstockFinance (2020), the bad debt ratio of Vietnamese commercial banks is all lower 3.42% in 2019 As a result, the imbalanced datasets effect on the efficiency of the good and bad classification algorithms at Vietnamese commercial banks This means that the number

of non-performing loans is much less than that of performing loans in the datasets These imbanced datasets can lead to biased prediction in favor of the major group in the classification al-gorithms (Ganganwar, 2012)

Reality shows the need to find algorithms for imbalanced data processing and debt clas-sification models appropriate to the context in Vietnam Earlier studies yet investigate the effec-tiveness of variable analysis based on grouped by WOE after managing imbalanced processed datasets This article aims to oversee an imbalanced credit dataset by oversampling, undersam-pling, bothsampling and SMOTE algorithms in the Vietnam bank context After that, we apply this handled imbalanced data into the logistic regression with input variables transformed by WOE to evaluate the importance of these algorithms and select an efficient credit scoring method

at commercial banks in Vietnam

2 Credit scoring method at commercial banks

The credit scoring process at commercial banks has a number of steps The first step is the data processing One of the challenging problems when applying regression models in practice

is the quality of the data The data processing step also consumes a great deal of time, accounting for nearly 80% of the total time The data processing includes segmenting, sampling, data parti-tioning, processing missing data, and outliers The next step is the scorecard construction process After data cleaning, the variables will be grouped by binning The new value for each group is WOE (Weight of Evidence) WOE presents the predictive power of an independent variable in relation to the dependent variable

Distribution of good is the percentage of good customers in a particular group

Distribution of bad is the percentage of bad customers in a particular group

The WOE approach is suitable for the logistic regression model thanks to the convenience

of scoring and no need to deal with missing data Binning is to reduce the number of attributes because if original variables are used in the regression model, a great deal of dummy variables will be created Also, binning variables is useful, especially for variables whose relationship to the dependent variable is nonlinear This technique helps deal with nonlinear effects in a linear model The next step is to remove some variables before running the model They are the variables which has the poor ability to differentiate between good and bad accounts Eliminating them be-fore running the model could improves the model quality and shortens processing time afterwards The index of value information – IV is used to remove weak variables (Siddiqi, 2012) The IV is calculated using the following formula:

Criteria for evaluating the variable predictiveness according to IV:

Table 1 The assessment of IV

(Source: Siddiqi, 2012)

< 0.02 Not useful for prediction 0.02 - 0.1 Weak predictive power 0.1 – 0.3 Medium predictive power – 0.5 Strong predictive power

>0.5 Suspicious predictive power

From the IV, some variables that are not useful for prediction or have weak predictive power are immediately removed from the model For remaining variables, it is necessary to com-bine with expert opinion to achieve the goal of dropping unnecessary variables but not losing good factors

Logit regression method

The dependent variable is the default status of the customer, symbolized by D (D = 0 if the customer pays on time, D = 1 if the customer pays late) Modelling the relationship between cus-tomer characteristics and repayment ability through the logit function as follows:

The maximum likelihood method is applied to estimate the coefficients β, thereby calcu-lating the default probability for each specific customer

Criteria for selecting variables for the logit model are statistical significance and expert judgement At the same time, operational efficiency evaluation includes considering resources mobilized for data collection, considering whether the use of variables is consistent with the pro-visions of law or not

3 Algorithms for handling imbalanced datasets

Oversampling

The common method of increasing instances in the group of bad accounts is randomly re-peating the data in this group Take a set of randomly selected minority examples in the minority group, then augment the size of this group by replicating the selected examples and adding them

to it The result is that total examples in the bad group will be increased and the class distribution

becomes more balanced accordingly

This random method simply replicates a portion of minority class in order to increase the weights of those examples Because the replacement process is totally random, this method re-creates some existing examples in the original minority class Therefore, its main drawback is that the overfitting phenomenon can occur This method is the most fundamental among over-sampling techniques Many other common overover-sampling algorithms used in the real world are developed based on this method (Peng Jun Huang, 2015)

Undersampling

Undersampling method randomly reduces observations in the good group The downside

of this technique is that it is possible to eliminate useful observations from the majority group The common technique applied in undersampling is the weighted sampling It is described

as follows: Calculate the weighted Euclidean distance of each negative instance from each of the positive instances All features are weighted by its Fishers discriminant score (F1 measure), which measures the overlapping per attribute After that, for each positive sample, sort negative instances

in ascending order of distance from the positive sample Finally, for each positive instance, select

a user-defined number of negative instances The user-defined number indicates the desired ratio

of negative samples to positive samples At this stage, special care is taken to avoid repetitive se-lection of negative samples If a particular negative sample has been already selected, the next available negative sample is selected (Fernández, 2018)

Bothsampling

Due to the fact that oversampling results in too many bad observations and undersampling can cause loss of the original data, the combination of these two methods can be chosen in the hope of fixing those problems However, it may not yield good results for the study either because the method simultaneously multiply bad observations and reduce the random good ones

SMOTE (Synthetic Minority Over-sampling Technique)

The SMOTE algorithm carries out an oversampling approach, but the key difference is that

it introduces synthetic examples but not replicate existing instances To create these new data points, the KNN algorithm is used for the minority group, clustering the data into K different groups Then, between two or more observations of the same group, it is needy to create more observations For this reason, the procedure is said to be focused on the “feature space” rather

than on the “data space” For example, an xi positive instance is selected as basis to create new synthetic data points Based on a distance metric, several NNs of the same class (points xi1 to

xi4) are chosen from the training set Finally, a randomized interpolation is carried out in order

to obtain new instances r1 to r4 (Fernández, 2018)

The SMOTE method works in “feature spaces” but not “data spaces” It treats nominal and continuous attributes in different ways In the closest neighbor determination calculations, it uses Euclidean distances for the continuum property and the Value Distance Metric for the nominal property

For continuum attributes:

hGet the difference between the feature vector (the minority class pattern) and one of its

k closest neighbors (the minority class sample)

hMultiply this difference with a random number between 0 and 1

hAdding this difference to the object value of the original object vector constitutes a new feature vector

For nominal attributes:

hTake the ratio of majority groups between the feature vector under consideration and the

k nearest neighbors In the case of a tie, take it at random

hAssign that value to an additional new case for the minority class

hUsing SMOTE creates additional regions for the minority, thereby allowing the classi-fication methods to predict more minorities

SMOTEboost algorithm

The SMOTEboost algorithm is the integration of the SMOTE algorithm into the standard

boosting process Thus, it gains the benefits of both boosting and SMOTE This algorithm applies

to some asymmetric data at the medium and high levels, improves the prediction results for the minority and improves the F-value

Boosting helps improve the predictability of classifiers by recalculating the weights of mis-classified observations SMOTE only improves the classifier of minority cases By attaching SMOTE to boosting, it promotes boosting that focuses more on minority cases than in the ma-jority

SMOTEboost implicitly increases the weight of false negatives in the distribution because the SMOTE algorithm increases the number of minority observations So, in the next iteration, SMOTEboost could create a wider decision area for the minority SMOTEboost combines SMOTE’s Recall-improving power and boosting’s Precision-improving power Altogether it im-proves the F-value or

4 Validation

In the classification problem, to evaluate the performance of the model, two methods are used including Confusion matrix and AUROC curve

Confusion Matrix

In these problems, it is common to define the more critical data class as the Positive class (P-positive), the other one called the Negative class (N-negative) In a good-bad classification problem, bad means Positive, and good means Negative From there, we define True Positive (TP), False Positive (FP), True Negative (TN), False Negative (FN) to create confusion matrix not normalized as follows:

Table 2 Confusion matrix

Table 3 Evaluation criteria

FPR (False Positive Rate) is also known as a false forecast rate, FNR (False Negative Rate)

is also known as omission rate

Predicted “bad” Predicted “good”

Actual “bad” True Positives (TP) False Negatives (FN) Actual “good” False Positives (FP) True Negatives (TN)

Evaluation criteria from the confusion matrix as following:

Predicted “bad” Predicted “good”

Actual “bad” TPR = TP/(TP+FN) FNR = FN/(TP+FN) Actual “good” FPR = FP/(FP+TN) TNR = TN/(FP+TN)

For the classification problem where the data sets of the classes are hugely different from each other, there is an efficiency measure commonly used as the Precision-Recall

Precision = TP/(TP+FP)

Recall = TP/(TP+FN) = TPR

High precision means the accuracy of the bad observations found is high Precision = 1 or

FP = 0 means that all observations predicted to be bad are true as bad However, this does not guarantee to find out all bad observations

A high recall means a high True Positive Rate, meaning a low rate of omitting bad obser-vations Recall = 1 or FN = 0 means finding all bad obserobser-vations However, it is unlikely that all bad prediction observations are correct

Therefore, a good classification model is one that has both Precision and Recall as high as possible, as close to 1 as possible To measure the quality of the classifier based on both Precision and Recall we use the F1score

Choosing β higher than 1 means to value Recall over Precision and vice versa, β less than 1 means that Precision is more important than Recall Two commonly used values of β are 2 and 0.5

AUROC curve

The values of FPR and TPR change when the good or bad threshold changes When the threshold decreases, both FPR and TPR increase, which means more false statements than omis-sions Conversely, when the threshold increases, both the FPR and TPR decrease, which means more omission than false statements

(Source: Sarang Narkhede, 2018)

Figure 1 Receiver Operating Characteristic curve or ROC curve

For each threshold, there is one pair of value (FPR, TPR) Standing for the points (FPR, TPR) on the graph, when changing the threshold from 0 to 1, we get a line called the Receiver Operating Characteristic curve or ROC curve

Plot the graph with the horizontal axis FPR (some graphs denoted by 1 - Specificity) and the vertical axis as TPR (some graphs denoted as Sensitivity)

Based on the ROC curve, one can show whether a model is useful or not An efficient model has low FPR and high TPR, there exists a point on the ROC curve that is close to the point with the coordinate (0, 1) on the graph (upper left corner) The closer the curve is, the more effi-cient the model is

There is another parameter used to evaluate a model called Area Under the Curve or AUROC, which is the area under the ROC curve This value is a positive number less than or equal to 1 The larger the value, the better the model The meaning of AUROC criteria

Table 4 Assessment of model by AUROC

(Source: D’Agostino et al., 2013)

Gini coefficient

The Gini coefficient is calculated according to the formula (Schechtman, 2016) of

Gini = 2* AUROC – 1 This coefficient is also used in evaluating the relevance or significance level of models Its measured values are between 0 and 1, where a score of 1 means that the model is 100% accurate

in predicting the outcome The closer the Gini is to one, the better model is On the other hand,

a Gini score equal to 0 means the model is entirely inaccurate

2 Results of experimental research

The article uses credit dataset at a commercial bank in Vietnam, including 7052 observa-tions with 456 bad observaobserva-tions With the bad rate of 7.13%, the data surveyed is imbalanced be-tween the two groups of good observations and the number of bad ones The data relating to the information of the customers is encrypted After outliers processing, imbalanced data is handled

by methods of oversampling, undersampling, bothsampling and SMOTE After that, the processed datasets are divided into two datasets including training set (70%) and testing set (30%) With each generated dataset, run the logistic model with the original variables and WOE grouping variables for credit scoring The model evaluation results are as follows:

> 0.90 Excellent 0.80 - 0.90 Good 0.70 - 0.80 Fair 0.60 - 0.70 Poor 0.50 - 0.60 Fail

Table 5 Results of logistic models

(Source: author’s calculation)

Using the original imbalanced dataset, the logistic model has the AUROC of 0.5521 and the G coefficient of 0.1042 This result shows that the model is not able to distinguish between good and bad observation The low precision and high Recall coefficients (0.008 and 0.999, re-spectively) means the accuracy of the bad observations found is low As a result, predict results are biased towards the majority group

After processing with oversampling, undersampling, bothsampling, and SMOTE methods, the validation results of the logit regression models with the unbalanced processed data set were improved compared to the model [1] The results show that the imbalanced handling method in-creases the efficiency of the classification methods The AUROC of these logistic models increase greater than 0.5, showing that these models are capable of differentiation Besides, the under-sampling method shows better results than the other three methods because the coefficients of Gini, Precision, Recall, F_score are all higher However, the Gini coefficient is in the range of 0.2 - 0.3, showing that the classification ability of these models is still quite weak

Model Logistic

with Method AUROC Gini Precision Recall F_score

[1] Original

variables Original 0.5521 0.1042 0.008 0.999 0.023 [2] Original

variables Oversam-pling 0.6271 0.254 0.581 0.499 0.308 [3]

Undersam-pling 0.6089 0.3166 0.628 0.784 0.448 [4]

Bothsam-pling 0.6296 0.2646 0.595 0.627 0.372 [5] SMOTE 0.6105 0.221 0.628 0.580 0.353

[6] (WOE)

grouping

variables

Oversam-pling 0.7051 0.4196 0.656 0.639 0.385

[7]

Undersam-pling 0.6228 0.244 0.599 0.634 0.376 [8]

Bothsam-pling 0.6978 0.3424 0.636 0.618 0.373 [9] SMOTE 0.8208 0.642 0.757 0.742 0.447