PREDICTING CLASS-IMBALANCED BUSINESS RISK USING RESAMPLING, REGULARIZATION, AND MODEL EMSEMBLING ALGORITHMS

International Journal of Managing Information Technology (IJMIT) Vol.11, No.1, February 2019
DOI : 10.5121/ijmit.2019.11101 1
PREDICTING CLASS-IMBALANCED BUSINESS RISK
USING RESAMPLING, REGULARIZATION, AND
MODEL EMSEMBLING ALGORITHMS
Yan Wang1
, Xuelei Sherry Ni2
1
Graduate College, Kennesaw State University, Kennesaw, USA
2
Department of Statistics and Analytical Sciences, Kennesaw State University,
Kennesaw, USA
ABSTRACT
We aim at developing and improving the imbalanced business risk modeling via jointly using proper
evaluation criteria, resampling, cross-validation, classifier regularization, and ensembling techniques.
Area Under the Receiver Operating Characteristic Curve (AUC of ROC) is used for model comparison
based on 10-fold cross validation. Two undersampling strategies including random undersampling (RUS)
and cluster centroid undersampling (CCUS), as well as two oversampling methods including random
oversampling (ROS) and Synthetic Minority Oversampling Technique (SMOTE), are applied. Three highly
interpretable classifiers, including logistic regression without regularization (LR), L1-regularized LR
(L1LR), and decision tree (DT) are implemented. Two ensembling techniques, including Bagging and
Boosting, are applied on the DT classifier for further model improvement. The results show that, Boosting
on DT by using the oversampled data containing 50% positives via SMOTE is the optimal model and it can
achieve AUC, recall, and F1 score valued 0.8633, 0.9260, and 0.8907, respectively.
KEYWORDS
Imbalance, resampling, regularization, ensemble, risk modeling
1. INTRODUCTION
Risk modeling can discriminate the risky business from the non-risky companies, thus
can guide the financial institutions to make decisions when processing loan or credit
applications [1] [2]. Logistic regression (LR) is a frequently used technique for risk
classifications since it is conceptually simple and explainable and has been demonstrated
to be powerful in many studies [3] [4] [5]. Decision tree (DT) is also widely used as it
has strong interpretability and relatively straightforward structures compared to
complicated models such as neural networks [6].
However, in real-life applications of risk models, many data are imbalanced. That is, the
distributions in each class are not uniform [7]. Imbalanced data result in several
problems and challenges to existing algorithms that have been shown to be effective on
the balanced data [8] [9] [10]. The problems arising from modeling imbalanced data
mainly contain the following three categories: (1) the usage of improper model
evaluation metric; (2) data rarity due to the lack of observations in the rare class; and (3)
the usage of weak classifiers or classifiers without regularization [11] [12] [13]. To
handle the problems mentioned above, researchers have proposed several effective
methodogies, including using more appropriate evaluate metrics, resampling
(oversampling or undersampling) with different ratios between positives and negatives,

2
using cross validation in the right way, classifier regularization such as L1-regularized
Logistic Regression (L1LR) and model ensembling [14] [15] [16] [17].
Motivated by previous research, we design a modeling work flow in this study, aiming at
developing a good model for imbalanced risk classifications by using the aforementioned
class-imbalance-handling methods simultaneously. The utilization of the above-
mentioned algorithms jointly is a major strength of this study, since there are few
previous research focusing on using these techniques in one modeling design. The
details of the modeling flow are described in Figure 3.
This paper is structured as follows. The relevant techniques used for handling class
imbalance are reviewed in Section 2. Section 3 describes the experimental design.
Experimental results are discussed in Section 4. Section 5 addresses the conclusions.
2. METHODOLOGIES FOR HANDLING CLASS IMBALANCE
In this section, the methodologies used in this study for classifying imbalanced targets
are reviewed.
2.1. Selection of the proper evaluation metrics
Classification accuracy is generally considered the optimal measure for binary classification.
However, it is no longer a good evaluation metric for models built on imbalanced data
because accuracy will bias towards the majority class. Instead, Area Under the Receiver
Operating Characteristic Curve (AUC of ROC) is more appropriate since it does not place
more emphasis on one class over another and it is not biased against the minority class [7]
[18] [19]. Moreover, ROC curve is independent of the positive-negative ratio in target,
making AUC a suitable measure for comparing different classifiers when the positive-
negative ratio varies [16]. Let P and N represent the total number of positive cases (i.e., risky
business) and negatives cases (i.e., non-risky business), respectively. Let True Positive (TP)
and False Positive (FP) denote those identified as risky business correctly or wrongly,
respectively. Similarly, we denote True Negative (TN) and False Negative (FN) as those
identified as non-risky business correctly or wrongly, respectively. Then recall (shown in
Equation 1) measures the fraction of risky business correctly classified and precision
(described in Equation 2) measures the fraction of objects classified as risky business that are
truly risky. Considering that a FN error may signify the loss, recall is used as the secondary
evaluation criterion in this study and precision is weighted less. Besides, F1 score (described
in Equation 3) is utilized as the third evaluation metric as it represents the harmonic mean of
precision and recall [20].
Recall =
𝑇𝑃
𝑇𝑃 + 𝐹𝑁
(1)
Precision =
𝑇𝑃
𝑇𝑃 + 𝐹𝑃
(2)
F1 score =
2 ∗ 𝑝𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 ∗ 𝑟𝑒𝑐𝑎𝑙𝑙
𝑝𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 + 𝑟𝑒𝑐𝑎𝑙𝑙
(3)

3
2.2. Resampling methods
Resampling, aiming at eliminating or minimizing the rarity by altering the distribution of
training samples, is a widely used while effective technique in the rare event prediction [21]
[22]. The resampling technique mainly includes two approaches: undersampling and
oversampling. The undersampling means removing observations in the majority class while
in oversampling, we duplicate the records in the minority class [18]. Figure 1 illustrates the
graphical representations of the above-mentioned two approaches. Both could balance the
class distribution, thus make the binary classifiers more robust. Although undersampling is
often used when the size of data is sufficient while oversampling is preferred when the data
is relatively small, there is no absolute advantage of one resampling method over another.
Application of oversampling or undersampling algorithms depends on different research
cases and dataset used [23].
Figure 1. Graphical representation of the undersampling and oversampling algorithms
In this paper, two undersampling techniques -- random undersampling (RUS) and cluster
centroid undersampling (CCUS) -- are implemented. RUS eliminates the examples in the
majority class (usually defined as the negatives in risk modeling) randomly and uniformly
while keeping all the observations from the minority class (usually defined as the positives in
risk modeling) to reach a desired positive percentage [23]. In CCUS, all observations in the
minority class of the original data are kept. Then, the clusters of the majority class are found
using K-means algorithms and only the K cluster centroids are kept for replacing the points
from the majority class [24]. The K in CCUS is set to the value that can reach the desired
percentage of positives.
In addition, two oversampling methods -- random oversampling (ROS) and Synthetic
Minority Oversampling Technique (SMOTE) -- are applied. Based on the similar idea as in
RUS, in ROS, the observations in the minority class are duplicated randomly and uniformly
while all the examples from the majority class are kept to reach the desired positive ratio
[25]. On the other hand, SMOTE does not duplicate the records in the minority class.
Instead, SMOTE could add newly synthesized instances (belonging to the minority class) to
the original data. And these newly generated examples should be similar to the positives
from the original data with respect to Euclidean distance. Figure 2 shows the details of the
SMOTE algorithm. In subplot (1) of Figure 2, the originally imbalanced data is represented
with a majority of negatives (blue points) and a minority of positives (orange points). In
subplot (2), we randomly select a positive instance (green point) and then find its k-nearest
Positives (P)Negatives (N)
Original imbalanced data with positive percentage P/(P+N)
Positives (P)Negatives (N’) Positives (P’)Negatives (N)
Eliminates N to N’ to reach
the desired positive
percentage P/(P+N’)
Duplicates P to P’ to reach the desired positive
percentage P’/(P’+N)
Undersample Oversample

4
neighbours among all the positives. In our study, we set k to be five and these neighbours are
denoted by the block dots marked with numbers. Next, in subplot (3), among the five
neighbours, one neighbour is randomly selected. The figure shows the illustrative case when
neighbour numbered 3 is selected. Then a new data point (denoted by the red dot) is
generated along the straight line connecting the neighbour numbered 3 and the green point.
Finally, this newly synthesized instance is labelled as positive and added to the original
dataset [26]. The procedure repeated for the positives to get the data with the desired positive
percentage.
2.3. Resampling to reach different positive percentages
During the resampling algorithm, one hyper-parameter we need to tune is the ratio between
positives and negatives (or, equivalently, the positive percentage in the training data). It is
shown by previous research that the best ratio (or, the best positive percentage) heavily
depends on the data and the models that are used, making it worth trying different
percentages and selecting the optimal one rather than training models using the same positive
percentage [27]. In this study, we use the resampling methods to obtain a series of training
datasets containing the positive percentages ranging from 10% to 90% with a step of 10%.
2.4. Using cross-validation in the proper way
Similar as comparing performance of classifiers on the balanced data, k-fold cross validation
is also frequently used when evaluating performance of models on class-imbalanced data.
The most important thing worthy to be emphasized is that k-fold cross validation should
always be done before, rather than after performing resampling strategies on the training data
(see details in Figure 3). If k-fold cross validation is applied after resampling, it will cause
the overlap of the training sets as well as leading to the over-fitting issue. On the contrary, if
we perform k-fold cross-validation before resampling, randomness can be introduced into the
dataset and the over-fitting problem can be reduced [28]. In our study, we set the value of k
to be 10 since many studies have shown that 10-fold cross validation can obtain efficiently
computational time while good model performance.
2.5. Regularization on logistic regression
It is shown that classic LR is not so robust without weighting and regularization when
being used on rare event data [29]. L1-regularized LR (L1LR), which is one of the
widely used regularized versions of LR, has been shown to outperform the classic LR in
modeling imbalanced data and high-dimensional data [1] [17] [30] [31].
L1LR works as follows. Suppose we have a training set with X∈Rn×d
, where n is the number
of observations and d denotes the number of features. For every instance 𝒙𝒊 ∈Rd
, the
dependent variable 𝑦𝑖 follows a Bernoulli distribution with probability of 𝑝𝑖 being 1 and
probability of (1-𝑝𝑖) being 0. LR models the relationship between each instance 𝒙𝒊 with its
expected outcome 𝑝𝑖 using Equation 4, where ( 𝛽0, 𝛽1, … , 𝛽 𝑛) is the parameter vector
and can be denoted by a vector β. By adding a regularization (penalty) term λ||β||1 into
the objective function (i.e., the log-likelihood function) of LR, we get the optimization
problem for L1LR defined in Equation 5, where λ is a hyper-parameter to control the
regularization and ||β||1 is the L1 norm of the parameters.

5
Figure 2. Graphical representation of the SMOTE algorithm
𝑝𝑖 =
exp(𝒙𝒊 𝛃)
1 + exp(𝒙𝒊 𝛃)
(4)
𝑚𝑖𝑛 ∑ −log 𝑝𝑖 𝑙(𝑦𝑖|𝒙𝒊; 𝛃
𝑛
𝑖=1
) + λ||𝛃|| 𝟏
(5)
2.6. Decision tree and model ensembling
Some studies point out that one of the reasons DT is not efficient for imbalanced data is,
in the building of DT, the instance space is partitioned into smaller and smaller spaces,
making it difficult to find the regularities [16]. However, there are still several studies
showing the promising utilization and the robustness of DT when being used on the
class-imbalanced data [32] [33]. Considering that DT is widely accepted in the financial
domain because of its strong interpretability as well as its robustness in the above-
mentioned studies, in our experiment we use it as one of the classifiers, aiming to
investigate whether it is still powerful in class-imbalanced risk classification. For the DT
algorithm, it searches the optimal splits on input variables based on different criteria
such as entropy or Gini index. Gini index is used in this paper and its calculation for a
given node is defined in Equation 6, where C is the number of classes in the dependent
variable and p(code) is the relative frequency of class c at the node [34].
𝐺𝑖𝑛𝑖( 𝑛𝑜𝑑𝑒) = 1 − ∑[𝑝(𝑐|𝑛𝑜𝑑𝑒)]2
𝐶
𝑐=1
(6)
Many studies have confirmed that comparing to a single learning algorithm, ensemble
methods can obtain better performance in rare event predictions by using multiple base
1. The original imbalanced data
containing positives (orange
dots) and negatives (blue dots).
2. Randomly pick a positive
instance (green dot). Then find its 5-
nearest neighbors and label them
(black dots).
3. Randomly pick a nearest
neighbor (neighbor 3). Then
synthesize a new positive instance
(red dot).
1
2
3
4
5
1
2
3
4
5

6
learners [35] [36]. Two widely used ensembling approaches are Bagging and Boosting.
Bagging builds a series of parallel while independent base classifiers using bootstrap
subsets of the training set. On the contrary, Boosting builds a series of base classifiers
sequentially and dependently. As a result, the subset used for building the current base
classifier is not generated by bootstrap sampling. Instead, the algorithm changes the
instance weights for the current classifier based on the performance of the previous
classifiers. Instances that are misclassified by previous classifiers would be weighted
more in the current classifier [37] [38]. In the end, both methods use a majority voting
logic for the final prediction. In this study, we consider both Bagging and Boosting
approaches based on the DT classifier, aiming to investigate whether we can obtain a
better risk model through ensembling. It is worth noting that before performing model
ensembling, we focus on getting a DT model as accurate as possible by identifying the
proper resampling method along with a properly resampled positive ratio, as described in
Sections 2.2 and 2.3. During the processing of Bagging and Boosting on DT classifier,
Gini index is also used to look for candidate variables for splitting each node in each DT
model. Moreover, Gini reduction is used to rank the variable importance for business
risk after the optimal model is identified.
3. EXPERIMENTAL DESIGN
3.1. Data description and pre-processing
The dataset provided by the US national Credit Bureau is used in our study to develop and
evaluate the risk models. The data contains commercial information of over 10 million de-
identified U.S. companies from 2006 to 2014. Example commercial information includes
business location, business size, liens, industry account activities, and liabilities. The
bankruptcy indicator represents the status of the business: 0 denotes the business is still
running while 1 means the business went to bankruptcy. To evaluate the risk of going
bankruptcy of the companies in 2014, a new target variable RiskInd is calculated based on
the change of the bankruptcy indicator. In other words, in our study, the positives (i.e., risky
business) are businesses which have changed the value of bankruptcy indicator from 0 in
2013 to 1 in 2014 while the negatives (i.e., non-risky business) are those having bankruptcy
indicator valued 0 in both 2013 and 2014. Observations that have bankruptcy indicator
valued 1 in both 2013 and 2014 are excluded in our modeling.
A series of data pre-processing procedures were applied sequentially as follows: (1) We
performed 10-fold cross validation in our study. In each of the 10-fold cross validation, 90%
data (i.e., 18,000 observations) was used as the training set while 10% (i.e., 2,000 records)
was used as the validation set; (2) Variables with more than 70% missing were removed; (3)
For both training and validation sets, missing values were imputed by the median of the
variable from the training set; (4) For both training and validation sets, variables were
standardized by using the means and the standard deviations of the variable from the training
set. The percentage of positives is about 7.4% and 36 independent variables are kept for
further analysis.
3.2. Methodology
Figure 3 shows the workflow of our numerical experiments by using oversampling as the illustrative
example. There are mainly five stages in the workflow. In stage 1, the experiment starts by
investigating the effect of different resampling methods on different classifiers. As described in
Section 2.3, in the oversampling experiments, ROS and SMOTE are implemented on the training
data to reach a series of resampled sets with different positive percentages: 10% to 90% with a
step of 10%. The same procedure is applied in the undersampling experiments, except that RUS

7
and CCUS are implemented. In stage 2, three classifiers including LR, L1LR, and DT are applied
on each of the resampled training sets. Then in stage 3, the model performance is evaluated and
compared on the validation set using AUC, recall, and F1 score. Stages 1, 2, and 3 are repeated 10
times in 10-fold cross validation and the optimal resampling method as well as the best resampled
positive percentage are selected for each of the three classifiers based on the average cross-
validated results. Therefore, in this stage, the most accurate base classifier is built. Next, in stage
4, Bagging and Boosting are implemented on DT using the training data with the optimal positive
percentage and the best resampling method from stage 3. The purpose is to examine whether the
performance of an individual DT can be further improved by model ensembling. Finally, in stage
5, the optimal model is selected and the important features related to risk classifications are
identified.
Figure 3. The workflow of the experimental design by using oversampling as the illustrative
example
4. EXPERIMENTAL RESULTS
With respect to the analysis tools, we use SAS (version 9.4) for the data pre-processing. Python
(version 3.5) is used for the rest of our analysis including resampling, 10-fold cross validation,
implementation of LR, L1LR, and DT, as well as Bagging and Boosting on DT. Our experiments
are implemented using a local desktop configured with a 3.3 GHz Intel Core i7 processor, 16GB
RAM, and MacOS system.
We started our experiments by first investigating the effect of different resampling methods on
model performance, and the results came out in the end of stage 3, as described in Figure 3.
Taking the 50% positive percentage value as an illustrative example, Figure 4 displays the
visualization of the resampled data with the first two principal components through principal
component analysis (PCA). The figure clearly shows that the sizes of the data obtained by RUS
and CCUS are obviously smaller than those obtained by ROS and SMOTE. Although all the four
resampled sets have 50% positives, the distributions of the positives and the negatives are
3. Scoring and evaluation, then
select the best positive percentage
and the best resampling method
for each classifier
Training set with
positive percentage
7.4% (blue denotes
positives while
yellow denotes
negatives) 1. Oversampling
on the training set
to reach different
positive
percentages
Training set LR
L1LR
DT
Boosting on DT
Bagging on DT
Validation set
2. Modeling on
each oversampled
data
10-fold cross validation
4. Model ensembling
Optimal model
Rank variables
5. Identify important
variables

8
different. The main difference is that the data obtained from SMOTE contains many newly
generated points while the rest three samples only contain the points from the original data. It is
because SMOTE algorithm synthesizes new points as illustrated by Figure 2.
Figure 4. PCA visualization of the training set with 50% positives obtained from four resampling
methods: RUS, CCUS, ROS, and SMOTE
In the end of stage 3 in the workflow, the average AUC of each classifier, LR, L1LR, or DT,
is calculated from 10-fold cross validation and the results are illustrated by Figures 5, 6, and
7, respectively. Overall, we see that, the different positive percentages result in very different
model performance. Furthermore, different resampling methods affect the model
performance in a various extent. In general, SMOTE outperforms the rest three resampling
methods on all three classifiers across almost all training sets with different positive
percentages. The highest AUC is obtained by SMOTE when the training set has 20%, 20%,
and 50% risky business on LR, L1LR, and DT, respectively (as marked by a circle in Figures
5, 6, and 7). We label the corresponding models as LR_SMOTE_20%, L1LR_SMOTE_20%,
and DT_SMOTE_50%, respectively.

9
Figure 5. AUC of LR versus various positive rates in the training data across different resampling methods
Figure 6. AUC of L1LR versus various positive rates in the training data across different resampling
methods

10
Figure 7. AUC of DT versus various positive rates in the training data across different resampling methods
To investigate the three models including LR_SMOTE_20%, L1LR_SMOTE_20%, and
DT_SMOTE_50% in more detail, we compare their performance using more evaluation
criteria described in Section 2.1. Furthermore, to identify the benefits of resampling, we have
implemented LR, L1LR, and DT on the original data (i.e., percentage of positives is 7.4%)
and the resulting models are labelled as LR_original_7.4%, L1LR_original_7.4%, and
DT_original_7.4%, respectively. All the above-mentioned results are illustrated in Figure 8.
By comparing performance between LR_original_7.4% and LR_SMOTE_20%, between
L1LR_original_7.4% and L1LR_SMOTE_20%, and between DT_original_7.4% and
DT_SMOTE_50%, we found that resampling strategy can improve AUC, recall, and F1
score in all these three classifiers, with recall having the most obvious increase. Surprisingly,
LR and L1LR have very similar performance by using all the evaluation criteria, no matter
whether the training data is resampled or not. This indicates that the LR with L1-
regularization does not show its superiority over LR without regularization in our study.
Figure 8. Comparison across different base classifiers and ensemble modelling using AUC, recall, and F1
score
AUC Recall F1 score
LR _orginal_7.4% 0.7456 0.8153 0.8477
L1LR_original_7.4% 0.7455 0.8149 0.8475
DT_original_7.4% 0.7511 0.7066 0.7763
LR_SMOTE_20% 0.7771 0.9130 0.8878
L1LR_SMOTE_20% 0.7771 0.9130 0.8878
DT_SMOTE_50% 0.7925 0.8003 0.8431
DT_SMOTE_50%_bagging 0.8645 0.7879 0.8353
DT_SMOTE_50%_boosting 0.8633 0.9260 0.8907
0.00
0.10
0.20
0.30
0.40
0.50
0.60
0.70
0.80
0.90
1.00
Modelperformance
Evaluations of different classifiers

11
In the stage 4 illustrated by Figure 3, two model ensembling techniques, Bagging and
Boosting, are applied on DT_SMOTE_50%, the DT classifier built on the optimal positive
rate as well as the best resampling method. The resulting models are labelled as
DT_SMOTE_50%_bagging and DT_SMOTE_50%_boosting, respectively. To make the
comparison of all the established models easier, the performance of
DT_SMOTE_50%_bagging and DT_SMOTE_50%_boosting is also illustrated in Figure 8.
It is observed that comparing with individual DT, while Bagging on DT improves AUC, it
hurts recall and F1 score, quite surprisingly. Compared to the base classifier DT, Boosting is
beneficial when considering AUC, recall, and F1 score. Moreover, Boosting on DT wins
Bagging by a huge margin in recall and F1 score, while Bagging outperforms Boosting by a
margin in terms of AUC. In the stage 5 illustrated by Figure 3, by combining all the
aforementioned discussions, we conclude that Boosting on DT by using the resampled
training set with 50% positives via SMOTE method (i.e., DT_SMOTE_50%_boosting) is the
optimal risk model in our study and it is labelled with a red star in Figure 8. It can achieve
AUC, recall, and F1 score valued 0.8633, 0.9260, and 0.8907, respectively.
Figure 9 shows the variable importance on business risk by using Gini reduction in the
DT_SMOTE_50%_boosting model. Higher score denotes more importance of the variable in
classifying the business risk. Results show that three variables including
MonLstRptDatePlcRec (i.e., number of months since the last report date on public records),
NocurLiensJud (i.e., number of current liens or judgment), and MostRecentLienorJud (i.e.,
most recent lien or judgment) have the largest weight comparing with other variables. On the
contrary, the variable such as pctSasNFA12mon (i.e., percentage of satisfactory non-
financial accounts in the last 12 months) shows a negligible effect in the risk classifications.
By ranking the risk factors, we can provide a comprehensive explanation on the possible
reasons of business risk occurrence.
Figure 9. Identified business risk factors ranked by Gini reduction in the optimal model

12
5. CONCLUSION
In business risk classification tasks, the data used is often imbalanced and it arises many
problems. In this study, we investigate how we can get a good risk model by simultaneously
using a series of algorithms, including using proper model evaluation metrics, resampling
along with different positive ratios, appropriate cross-validation, classifier regularization and
ensembling. The simultaneous utilization of the above-mentioned algorithms is a major
strength of this study, as there are few previous studies considering these techniques jointly.
In our experiment, AUC, rather than the widely-used accuracy, is selected as the primary
measure when comparing different models. Two undersampling methods (RUS and CCUS)
and two oversampling methods (ROS and SMOTE) are applied to resample the training data
to reach a series of positive ratios ranging from 10% to 90% using a step of 10%. Three
classifiers with strong interpretations including LR, L1LR, and DT are implemented on the
resampled data as well as on the original data. Their performance is compared through 10-
fold cross validation. Two model ensembling techniques, including Bagging and Boosting,
are used on DT to further improve the DT performance. Finally, important features related to
business risk are identified by using the optimal model.
Compared to the models built on the data without resampling, we found the increase in AUC,
recall, and F1 score in all the three classifiers including LR, L1LR, and DT after resampling
the training set to an appropriate positive percentage. SMOTE is shown to be the best
resampling method for LR, L1LR, and DT across each percentage of positives in the training
set. By using SMOTE, LR, L1LR, and DT can reach the highest AUC by using the
resampled data with 20%, 20%, and 50% positives, respectively. On the other hand, among
the four resampling methods, CCUS results in the lowest AUC in LR and DT when
percentage of positives is between 20% and 80%. RUS and CCUS produce obviously
different AUC on LR and DT. Although RUS and CCUS produce similar performance in
L1LR, CCUS is less preferred because of its long processing time. Surprisingly, L1LR did
not outperform LR in our result, since they produce very similar AUC under the same
resampling procedure. By comparing with the base DT classifier, Boosting on DT is
beneficial in terms of AUC, recall, and F1 score while Bagging on DT only improves AUC.
The unexpected outcome is that Bagging on DT even slightly decrease recall and F1 score
comparing to individual DT. The optimal candidate model for risk modeling is the Boosting
model based on DT by using the resampled data with 50% positives via SMOTE. It can
achieve the AUC, recall, and F1 score valued 0.8633, 0.9260, and 0.8907, respectively.
MonLstRptDatePlcRec is shown to be the most important feature in classifying business risk
while pctSasNFA12mon has little predictive power.
There is no general answer of the best resampling methods along with the resampled positive
percentage in building risk models, as the answer is surely case and data dependent.
However, the experimental design used in this study can serve as a reference for future
studies in developing and improving risk models on imbalanced data. Moreover, the critical
variables for the business risk classification provided by the Boosting DT model in our study
could guide financial institutions in approving loan applications.
REFERENCES
[1] R. Calabrese, G. Marra, and S. Angela Osmetti, “Bankruptcy prediction of small and medium
enterprises using a flexible binary generalized extreme value model,” Journal of the Operational
Research Society, vol. 67, no. 4, pp. 604–615, 2016.

13
[2] J. Galindo and P. Tamayo, “Credit risk assessment using statistical and machine learning: basic
methodology and risk modeling applications,” Computational Economics, vol. 15, no. 1-2, pp.
107–143, 2000.
[3] N. Scott G. Sunil, and K. Wagner. “Defection detection: Measuring and understanding the
predictive accuracy of customer churn models,” Journal of marketing research, vol.43, no. 2, pp.
204-211, 2006.
[4] Y. Wang, X. S. Ni, and B. Stone, “A two-stage hybrid model by using artificial neural networks as
feature construction algorithms,” Internal Journal of Data Mining & Knowledge Management
Process (IJDKP), vol. 8, no. 6, 2018.
[5] B. Baesens, T. Van Gestel, S. Viaene, M. Stepanova, J. Suykens, and J. Vanthienen,
“Benchmarking state-of-the-art classification algorithms for credit scoring,” Journal of the
operational research society, vol. 54, no. 6, pp. 627–635, 2003.
[6] M. Zekic-Susac, N. Sarlija, and M. Bensic, “Small business credit scoring: a comparison of
logistic regression, neural network, and decision tree models,” Information Technology Interfaces,
2004. 26th International Conference on IEEE, 2004, pp. 265–270.
[7] N. V. Chawla, N. Japkowicz, and A. Kotcz, “Special issue on learning from imbalanced data sets,”
ACM Sigkdd Explorations Newsletter, vol. 6, no. 1, pp. 1–6, 2004.
[8] Y. Zhou, M. Han, L. Liu, J. S. He, and Y. Wang, “Deep learning approach for cyberattack
detection,” IEEE INFOCOM 2018-IEEE Conference on Computer Communications Workshops
(INFOCOM WKSHPS). IEEE, 2018, pp. 262–267.
[9] G. King and L. Zeng, “Logistic regression in rare events data,” Political analysis, vol. 9, no. 2, pp.
137–163, 2001.
[10] N. Japkowicz and S. Stephen, “The class imbalance problem: A systematic study,” Intelligent data
analysis, vol. 6, no. 5, pp. 429–449, 2002.
[11] M. Schubach, M. Re, P. N. Robinson, and G. Valentini, “Imbalance-aware machine learning for
predicting rare and common disease-associated non-coding variants,” Scientific reports, vol. 7, no.
1, p. 2959, 2017.
[12] V. Lopez, A. Fernandez, S. Garcia, V. Palade, and F. Herrera, “An insight into
classification with imbalanced data: Empirical results and current trends on using data intrinsic
characteristics,” Information Sciences, vol. 250, pp. 113–141, 2013.
[13] M. Galar, A. Fernandez, E. Barrenechea, H. Bustince, and F. Herrera, “A review on ensembles for
the class imbalance problem: bagging-, boosting-, and hybrid- based approaches,” IEEE
Transactions on Systems, Man, and Cybernetics, Part C (Applications and Reviews), vol. 42, no.
4, pp. 463–484, 2012.
[14] Y. Xia, C. Liu, Y. Li, and N. Liu, “A boosted decision tree approach using bayesian hyper-
parameter optimization for credit scoring,” Expert Systems with Applications, vol. 78, pp. 225–
241, 2017.
[15] G. Wang and J. Ma, “A hybrid ensemble approach for enterprise credit risk assessment based on
support vector machine,” Expert Systems with Applications, vol. 39, no. 5, pp. 5325–5331, 2012.
[16] J. Burez and D. Van den Poel, “Handling class imbalance in customer churn prediction,” Expert
Systems with Applications, vol. 36, no. 3, pp. 4626–4636, 2009.

14
[17] D. Muchlinski, D. Siroky, J. He, and M. Kocher, “Comparing random forest with logistic
regression for predicting class-imbalanced civil war onset data,” Political Analysis, vol. 24, no. 1,
pp. 87–103, 2016.
[18] B. W. Yap, K. A. Rani, H. A. A. Rahman, S. Fong, Z. Khairudin, and N. N. Abdullah, “An
application of oversampling, undersampling, bagging and boosting in handling imbalanced
datasets,” Proceedings of the First International Conference on Advanced Data and Information
Engineering (DaEng-2013). Springer, 2014, pp. 13–22.
[19] N. Japkowicz et al., “Learning from imbalanced data sets: a comparison of various strategies,”
AAAI workshop on learning from imbalanced data sets, vol. 68. Menlo Park, CA, 2000, pp. 10–
15.
[20] C. Goutte and E. Gaussier, “A probabilistic interpretation of precision, recall and f-score, with
implication for evaluation,” European Conference on Information Retrieval. Springer, 2005, pp.
345–359.
[21] M. Pavlou, G. Ambler, S. R. Seaman, O. Guttmann, P. Elliott, M. King, and R. Z. Omar, “How to
develop a more accurate risk prediction model when there are few events,” Bmj, vol. 351, p.
h3868, 2015.
[22] Y. Zhang and P. Trubey, “Machine learning and sampling scheme: An empirical study of money
laundering detection,” 2018.
[23] N. V. Chawla, “Data mining for imbalanced datasets: An overview,” Data mining and knowledge
discovery handbook. Springer, 2009, pp. 875–886.
[24] S.-J. Yen and Y.-S. Lee, “Cluster-based under-sampling approaches for imbalanced data
distributions,” Expert Systems with Applications, vol. 36, no. 3, pp. 5718–5727, 2009.
[25] A. Liu, J. Ghosh, and C. E. Martin, “Generative oversampling for mining imbalanced datasets.”
DMIN, 2007, pp. 66–72.
[26] N. V. Chawla, K. W. Bowyer, L. O. Hall, and W. P. Kegelmeyer, “Smote: synthetic minority over-
sampling technique,” Journal of artificial intelligence research, vol. 16, pp. 321–357, 2002.
[27] L. Zhang, J. Priestley, and X. Ni, “Influence of the event rate on discrimination abilities of
bankruptcy prediction models,” Internal Journal of Database Management Systems (IJDMS), vol.
10, No. 1, 2018.
[28] L. Lusa et al., “Joint use of over-and under-sampling techniques and cross- validation for the
development and assessment of prediction models,” BMC bioinformatics, vol. 16, no. 1, p. 363,
2015.
[29] M. Maalouf and T. B. Trafalis, “Robust weighted kernel logistic regression in imbalanced and rare
events data,” Computational Statistics & Data Analysis, vol. 55, no. 1, pp. 168–183, 2011.
[30] M. Maalouf and M. Siddiqi, “Weighted logistic regression for large-scale imbalanced and rare
events data,” Knowledge-Based Systems, vol. 59, pp. 142–148, 2014.
[31] P. Ravikumar, M. J. Wainwright, J. D. Lafferty et al., “High-dimensional ising model selection
using 1-regularized logistic regression,” The Annals of Statistics, vol. 38, no. 3, pp. 1287–1319,
2010.
[32] H. Y. Chang, D. S. Nuyten, J. B. Sneddon, T. Hastie, R. Tibshirani, T. Sørlie, H. Dai, Y. D. He, L.
J. van’t Veer, H. Bartelink et al., “Robustness, scalability, and integration of a wound-response
gene expression signature in predicting breast cancer survival,” Proceedings of the National
Academy of Sciences of the United States of America, vol. 102, no. 10, pp. 3738–3743, 2005.

15
[33] W. Liu, S. Chawla, D. A. Cieslak, and N. V. Chawla, “A robust decision tree algorithm for
imbalanced data sets,” Proceedings of the 2010 SIAM International Conference on Data Mining.
SIAM, 2010, pp. 766–777.
[34] I. H. Witten, E. Frank, M. A. Hall, and C. J. Pal, Data Mining: Practical machine learning tools
and techniques. Morgan Kaufmann, 2016.
[35] S. Bhattacharyya, S. Jha, K. Tharakunnel, and J. C. Westland, “Data mining for credit card fraud:
A comparative study,” Decision Support Systems, vol. 50, no. 3, pp. 602–613, 2011.
[36] Y. Sun, M. S. Kamel, A. K. Wong, and Y. Wang, “Cost-sensitive boosting for classification of
imbalanced data,” Pattern Recognition, vol. 40, no. 12, pp. 3358– 3378, 2007.
[37] E. Bauer and R. Kohavi, “An empirical comparison of voting classification algorithms: Bagging,
boosting, and variants,” Machine learning, vol. 36, no. 1-2, pp. 105–139, 1999.
[38] Y. Wang and X. S. Ni, “A XGBoost risk model via feature selection and Bayesian hyper-
parameter optimization,” https://arxiv.org/abs/1901.08433.
AUTHORS
Yan Wang is a Ph.D. candidate in Analytics and Data Science at Kennesaw State
University. Her research interest contains algorithms and applications of data
mining and machine learning techniques in financial areas. She has been a summer
Data Scientist intern at Ernst & Young and focuses on the fraud detections using
machine learning techniques. Her current research is about exploring new
algorithms/models that integrates new machine learning tools into traditional
statistical methods, which aims at helping financial institutions make better
strategies. Yan received her M.S. in Statistics from university of Georgia.
Dr. Xuelei Sherry Ni is currently a Professor of Statistics and Interim Chair of
Department of Statistics and Analytical Sciences at Kennesaw State University,
where she has been teaching since 2006. She served as the program director for the
Master of Science in Applied Statistics program from 2014 to 2018, when she
focused on providing students an applied leaning experience using real-world
problems. Her articles have appeared in the Annals of Statistics, the Journal of
Statistical Planning and Inference and Statistica Sinica, among others. She is the also
the author of several book chapters on modeling and forecasting. Dr. Ni received
her M.S. and Ph.D. in Applied Statistics from Georgia Institute of Technology.

PREDICTING CLASS-IMBALANCED BUSINESS RISK USING RESAMPLING, REGULARIZATION, AND MODEL EMSEMBLING ALGORITHMS

More Related Content

What's hot (19)

Similar to PREDICTING CLASS-IMBALANCED BUSINESS RISK USING RESAMPLING, REGULARIZATION, AND MODEL EMSEMBLING ALGORITHMS (20)

More from IJMIT JOURNAL (20)

Recently uploaded (20)

PREDICTING CLASS-IMBALANCED BUSINESS RISK USING RESAMPLING, REGULARIZATION, AND MODEL EMSEMBLING ALGORITHMS