This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2019.2927602, IEEE
Access

Date of publication xxxx 00, 0000, date of current version xxxx 00, 0000.
Digital Object Identifier 10.1109/ACCESS.2017.Doi Number

A Novel Reject Inference Model Using Outlier

Detection and Gradient Boosting Technique in
Peer-to-Peer Lending
Yufei Xia1
1
Businees School, Jiangsu Normal University, Xuzhou, Jiangsu 221116 PR China

Corresponding author: Yufei Xia (e-mail: 6020180093@jsnu.edu.cn).

This work was supported in part by Research Support Project for Doctoral Degree Teachers of Jiangsu Normal University under Grant 18XWRX021.

ABSTRACT Credit scoring is an efficient tool in handling the information asymmetry of P2P lending.
Typically, credit scoring models are built only with the accepted applicants, which may cause sample bias
and further hinder the predictive performances. Reject inference methods utilize the information contained in
the rejected samples by inferring their potential status and incorporate them with the accepted samples. In
this paper, we propose a novel reject inference model (i.e., OD-LightGBM) that combines an outlier detection
technique (i.e., isolation forest) and a state-of-the-art gradient boosting decision tree algorithm. The model is
evaluated on two real-world P2P lending datasets and the results of predictive performances demonstrate that
our proposal significantly outperforms the benchmarks in terms of discriminative capability. The analysis of
computational cost shows that our proposal has a great potential in handling large-sized problems. The
proposed framework remains robust under different parameter settings and provides stable results given
various combinations of outlier detection algorithms and classifiers.

INDEX TERMS credit scoring, gradient boosting decision tree, outlier detection, P2P lending, reject
inference

I. INTRODUCTION lenders can hardly know the borrowers’ willingness and

FinTech, a close integration of IT and financial sectors, is
emerging rapidly across the world. According to KMPG,
global FinTech investment doubled in 2018, reaching to USD
111.8 million1. Among areas of FinTech, Peer-to-Peer lending
(also known as social lending) has received much attention in
China, partially due to its critical role of inclusive finance [1].
In P2P lending, borrowers and lenders are matched directly via
online platforms. The platforms function as information
intermediaries that transfer concerns between borrowers and
lenders. P2P lending is typically operated online and bypasses
bank. Thus, P2P lending is usually convenient and leads to
lower transaction costs. Despite the considerable benefits, P2P
lending is characterized as inherent high risk due to lack of
collateral and information asymmetry [2, 3]. Specifically, the
lenders may suffer from huge loss due to the non-performing
P2P loans when collateral is limited. Moreover, the borrowers
have information advantages over lenders. In other words, the
capabilities to pay. Thus, P2P lending is criticized as high
information asymmetry [4].
To avoid the market failure resulted by information
asymmetry [5], various methods have been employed in field
of P2P lending. In early period of P2P lending, group lending,
a concept derived from microfinance, is a popular way to deal
with information asymmetry. The members of group monitor
each other before extending credit and punish defaulters via
informal enforcement mechanisms after the loans are assigned.
By these means, group lending mitigates the adverse selection
and moral hazard [6]. Though Prosper, a popular P2P lending
platform, established lending group at its very beginning, such
a mechanism is gradually abandoned in recent P2P lending,
partially because group membership does not take any
collateral responsibility or make any interaction when the
credit is extended. Internal credit scoring is another important
tool to deal with information asymmetry in P2P lending.

1
https://home.kpmg/xx/en/home/insights/2019/01/pulse-of-fintech-h2-
2018.html

VOLUME XX, 2019

This work is licensed under a Creative Commons Attribution 3.0 License. For more information, see http://creativecommons.org/licenses/by/3.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2019.2927602, IEEE
Access
Yufei Xia: A Novel Reject Inference Model Using Outlier Detection and Gradient Boosting Technique in Peer-to-Peer Lending (June
2019)
Unskilled lenders may suffer from marked biases and the
prospective incapability to convert the information into
profitable decision-making. Therefore, the platforms usually
build credit scoring system to screen the potential borrowers.
Abundant research has demonstrated that credit scores affect
the decision-making of borrowers [7-9]. The rationale behind
credit scoring is to apply classification method based on the
application characteristics to predict the probability default
(PD) of loan requests. If the PD exceeds a preset threshold,
then the loan application would be rejected. The economic
benefits of credit scoring is also significant: even a minor
improvement of credit scoring system would prevent dramatic
loss for lenders [10].
Current studies mainly optimize the credit scoring
techniques from two aspects, namely the methodology and
data. Regarding the methodology, there is a growing trend that
AI-based techniques are within the mainstream of credit
scoring models instead of traditional statistical methods. As
the award-winning research by Lessmann, et al. [11] claimed,
several AI-based methods outperforms the industry
benchmark (i.e., logistic regression (LR)). Thus, ensemble
learning, deep learning and other state-of-the-art classifiers
have been applied to credit scoring. In terms of data, both
academia and industry explore predictive variables to further
enhance the performance of credit scoring. Soft information is
a valuable data source of credit scores recently. In a thorough
discussion of Liberti and Petersen [12], the authors highlighted
the role of soft information in credit risk assessment,
especially for P2P lending. Besides the hard and soft
information in accepted loans, rejected loan applications also
contain crucial information that potentially benefits credit
scoring. It is worth mentioning that most credit scoring models
are established on historical performances, which means that
only accepted requests have been used in estimating PDs.
However, the amount of rejected loan request is much larger
than the accepted ones especially in P2P lending: roughly 9%
of loan applications are accepted and the remaining over 90%
of loans are rejected [13]. Reject inference, therefore, a
process makes educated guesses on how rejected loans would
have performed if accepted, is an efficient way to improve data
size in credit scoring.
The main contribution of this study includes the following
two aspects. First, we develop a novel reject inference
framework (OD-LightGBM) that combines a recent outlier
detection algorithm (i.e., isolation forest) and state-of-the-art
GBDT classifier (LightGBM). Several studies have applied
unsupervised outlier detection in credit scoring as a pre-
processing step for the purpose of mitigating negative effect
of outliers in training classifiers [14-16]. To the best of our
knowledge, there is no existing study that employs outlier
detection algorithm in inferring the potential status of rejected
requests. Moreover, the superiority of isolation forest
algorithm is empirically demonstrated over various tasks [17].
LightGBM is an efficient GBDT algorithm that achieves
better accuracy in many machine learning tasks [18].
This work is licensed under a Creative Commons Attribution 3.0 License. For more information, see http://creativecommons.org/licenses/by/3.0/.
Compared with other GBDT algorithms such as GBM and
XGBoost, LightGBM makes some modifications in training
base learners. Moreover, LightGBM supports GPU-
computing, which can deal with large-sized problems.
Consequently, the proposed reject inference framework is
expected to be high-performance, flexible, and scalable.
Second, we conduct a comprehensive sensitivity analysis on
reject inference. In most concerning studies, the proportion of
accepted and rejected loans (i.e., rejection rate) is fixed. Using
such a fixed dataset may lose generalization capability or
cause a waste on computing resources since the rejection rate
of P2P lending is extremely high. This paper addresses these
issues via sensitivity analysis. Concretely, we sample the
rejected dataset to generate multiple datasets with different
rejection rates. The proposed framework is validated on two
real-world datasets to examine whether rejection rate affects
the model performances. Moreover, the usage of outlier
detection algorithm provides us an opportunity to determine
the contamination rate (i.e., the proportion of outliers or Good-
Bad ratio, abbreviated as CR) in inference results. Thus, we
aim to verify the performances of our proposal under different
scenario and further present some recommendations on the
data pre-processing in practical application.
The rest of this paper is structured as follows. In Section II,
we conduct a comprehensive literature review concerning
reject inference in P2P lending. Section III introduces the
methodology. The mechanism of GBDT, LightGBM and the
proposed OD-LightGBM are demonstrated. Section IV shows
the data used in this paper. The experiment setup and results
are explained in Section V and finally in Section VI, main
conclusions and future research are presented.
II. LITERATURE REVIEW
A. REJECT INFERENECE IN P2P LENDING
Most of credit scoring models are built based on the
information of accepted loans which were previously
considered as trustworthy borrowers. However, the models
should be established on information of all requests
theoretically, otherwise it may cause sample bias [19]. Sample
bias is inherently a missing data dilemma that can be further
divided into missing at random (MAR) and missing not at
random (MNAR) [20]. The former indicates that the missing
data mechanism is independent of the loan characteristics (i.e.,
features) and the status (i.e., labels). In such a situation, the
parameters estimated from a credit scoring model using only
accepted requests do not exhibit sample bias. As indicated by
Feelders [21], expectation-maximization is an efficient model
to execute reject inference when missing mechanism is MAR
However, if the data is MNAR, which means that the rejected
outcomes are related with the features and labels, sample bias
would occur. Sample bias may result in biased parameters and
poor predictability [13].
Reject inference is regarded as a remedy to sample bias. A
variety of techniques have been applied in reject inference,
among which can be roughly categorized into two subsets, the
VOLUME XX, 2019
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2019.2927602, IEEE
Access
statistical methods and machine learning techniques. The early
reject inference research mainly considers statistical methods,
including augmentation, re-weighting, extrapolation and
survival analysis [19, 20, 22]. On contrast, recent machine
learning methods usually utilize semi-supervised SVM for
reject inference [13, 23]. Interestingly, Chen and Åstebro [24]
proposed a Bayesian analysis model which solved the missing
value issue in reject inference by a bound and collapse
imputation technique.
Many scholars have also demonstrated the efficiency of
internal credit scores in affecting lenders’ behavior and
identifying potential risky borrowers [8, 25]. Thus, the missing
mechanism in credit scoring in the specific P2P lending is also
likely to be MNAR since accepted loans are selected
according to the expected credit quality [24]. Such a missing
mechanism, along with the high rejection rate, highlights the
necessity of reject inference in P2P lending. Some studies
have developed SVM-based reject inference models for P2P
lending. Among them, Li, et al. [13] is a pioneering research,
which uses semi-supervised SVM (S3VM) for reject inference.
Validated on Lending Club dataset, the proposed S3VM
outperforms industry benchmark LR. Tian, et al. [16] further
develop a kernel-free fuzzy quadratic surface SVM reject
inference model, which addresses the issues of hyper-
parameter tuning and scalability. A recent study of Kim and
Cho [26] combines combine label propagation and semi-
supervised SVM to infer the potential status of reject
applicants. Besides SVM, GBDT is also applied in field of
reject inference as classifier. Xia, et al. [27] propose a reject
inference framework by integrating contrastive pessimistic
likelihood estimation and LightGBM. All the above studies
have shown the superiority of reject inference compared with
supervised credit scoring model. Table 1 gives a summarize of
the research on reject inference in P2P lending.
Insert Table 1 here
As indicated in Table 1, the rejection rate is usually fixed.
Semi-supervised SVM is typically considered as both reject
inference approach and classifiers, and few studies apply
ensemble classifiers in reject inference. On the contrary,
multiple datasets with different rejection rates are employed to
validate the proposed ensemble reject inference model in this
paper. The ensemble architecture is expected to improve the
performance of reject inference models and various rejection
rates would enhance the robustness of the proposal.
B. OUTLIER DETECTION IN CREDIT SCORING
An outlier indicates any data point that differs greatly from
other samples as to doubt it was generated by a different
mechanism [28]. An early application of outlier detection in
credit risk assessment would be the fraud detection [29]: the
fraudulent applications and risky customers may have
different patterns from normal ones, which is suitable for
VOLUME XX, 2019
This work is licensed under a Creative Commons Attribution 3.0 License. For more information, see http://creativecommons.org/licenses/by/3.0/.
usage of outlier detection. Some representative studies include
Srivastava, et al. [30], and Panigrahi, et al. [31].
Credit scoring is inherently a classification model. However,
some classifiers, such as SVM and decision tree, are very
sensitive to outliers and noisy data [32]. Outlier detection may
benefit supervised learning mainly from two aspects. First,
outlier detection reduces the data noise and retain clean data
[33]. Second, removing outliers can decrease the data size and
save computing resources [34]. Thus, some scholars have
attempted to apply outlier detection in pre-processing of credit
scoring. García, et al. [14] claim that the datasets that removes
outliers outperforms the original datasets significantly when
training credit scoring model. The empirical results of Setiono,
et al. [35] and Tian, et al. [16] also demonstrate the superiority
of outlier detection as pre-processing tool. However, few
studies have applied outlier detection in reject inference.
C. CREDIT SCORING IN P2P LENDING
Due to the critical role of credit scoring in P2P lending, a
variety of studies focused on the specific P2P lending domain.
These models are roughly categorized into the statistical
models and AI-based models. LR is one of the most popular
statistical models in P2P lending mainly due to its acceptable
performances and interpretability [36]. AI-based methods,
such as support vector machine [37], neural network [38], and
deep neural network [39] have been applied in credit scoring
of P2P lending due to their superior predictability [11, 40]. For
the pursuit of better predictability, ensemble learning is a hot
research field in terms of methodology of credit scoring.
Random forest is a representative model among ensemble
methods [7, 41] and achieves a balance between performance
and complexity. Recently, because of its sound performance,
gradient boosting decision tree (GBDT) have been widely
applied to credit risk assessment in P2P lending [42].
Specifically, Xia, et al. [36] developed a credit scoring based
on an advanced GBDT algorithm (i.e., XGBoost) to measure
PD of P2P lending applications. Ma, et al. [43] combined data
cleaning techniques and GBDT techniques in predicting loan
default. Furthermore, heterogeneous ensemble models, which
combine the predictions of different classifiers, are also
proposed by some scholars [42, 44].
III. METHODOLOGY
As indicated in Lessmann, et al. [11], ensemble credit scoring
models perform well and resisting them in practice is more
psychological than business related. The basic idea of
ensemble learning is to integrate multiple “weak” models into
a “strong” one. Bagging and boosting are two representative
types of ensemble learning. Boosting herein means iteratively
training weak models and then adding them to a final model.
For the specific gradient boosting, the base model is usually
decision tree. Hence, we usually call such a model as GBDT.
GBDT is a family of learning algorithms, including some
recent variations such as XGBoost and LightGBM. Thus, it is
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2019.2927602, IEEE
Yufei Xia: A Novel Reject Inference Model Using Outlier Detection and Gradient Boosting Technique in Peer-to-Peer Lending (June
2019)
necessary to provide an overview on GBDT before
introducing LightGBM.
A. AN OVERVIEW OF GBDT
GBDT builds models in a step-wise manner that optimizes an
arbitrary differentiable objective function [45]. Formally,
given a training set X = {(𝐱 , y )} , the target of GBDT is
to search for an approximation 𝐹 (𝒙) to a specific function
𝐹(𝒙) for the purpose of minimizing the expected value of a
certain objective function L(𝑦, 𝐹(𝒙)), namely,
𝐹 = argmin 𝐸 𝐿 𝑦, 𝐹(𝒙) .
, (1)
Since GBDT can handle regression, classification, and
ranking problems, the objective function may vary for
different learning tasks. For classification tasks, the
commonly-used objective function is log loss or hinge loss.
For details of the two objective functions, please see Bishop
[46]. As aforementioned, GBDT typically add the based
models together to form a new model, that is,
𝐹 (𝑥) = 𝑓 (𝑥),
where T is the number of base models. The subsequent issue
is to determine the optimal base model. The optimization
process at the t-th iteration can be described as
𝑓 (𝑥) = argmin ∈ ℒ
= argmin ∈ (𝐿(𝑦, 𝐹 (𝑥) (3)
+ 𝑓 (𝑥))).
The optimal base learner can be searched in different ways.
Classical GBDT technique applies the steepest descent
method to determine the optimal base model. Such a method
employs only the first-order gradient of the training sample
given an objective function. However, in state-of-the-art
GBDT techniques (e.g., XGBoost and LightGBM), the
objective function is quickly approximated via Newton–
Raphson algorithm. Note that for classification task, the base
model in GBDT is normally determined as classification and
regression tree (CART). A CART is a tree structure, which
splits the data samples into different regions (leaves) following
a series of “if-then-else” rules. Thus, CART is interpretable
and can be easily understood. Building a CART mainly
consists of three successive steps, namely selecting splitting
variable, determining cut-off value and pruning. Regarding
original GBDT, the base learners are trained in a level-wise
learning. All the variables and possible cut-off values are
evaluated using objective function and selected in a greed
manner. Once a leaf is decided, it stops further partition until
the leaves in the same level are determined (please see Fig.1
(a)). Let denote the base learner at t-th as follows
𝑓 (𝑥) = 𝑏 𝟏 (𝑥),
This work is licensed under a Creative Commons Attribution 3.0 License. For more information, see http://creativecommons.org/licenses/by/3.0/.
Access
where 𝐽 is the number of leaves. CART partitions the
samples into a number of 𝐽 disjoint regions and these regions
are denoted as 𝑅 . 𝑏 is the prediction of region 𝑅 . The
coefficients 𝑏 are usually multiplied by some parameters γ .
γ is determined in line search method, namely
𝛾 = 𝑎𝑟𝑔𝑚𝑖𝑛 𝛹(𝑦, 𝐹 (𝑥) + 𝛾𝑓 (𝑥)). (5)
GBDT is therefore updated as follows:
𝐹 (𝑥) = 𝐹 (𝑥) + 𝛾 𝑓 (𝑥) (6)
B. LIGHTGBM ALGORITHM
LightGBM is a powerful and distributed GBDT algorithm
proposed by Ke, et al. [18]. Such a high-performance
algorithm has become the winning solution of many machine
learning competitions. LightGBM technique can handle
various task, such as classification, regression, and ranking.
Several studies have demonstrated the superiority of
LightGBM in many applications, such as tumor classification
[47] and loan default prediction [43]. The promising
performances of LightGBM can be partially explained by the
application of leaf-wise learning. Different from GBDT,
(2) LightGBM builds best-first tree [48], which chooses the leaf
with expected best objective function to grow (see Fig. 1 (b)).
It is worth mentioning that there is a trade-off between level-
wise and leaf-wise trees. The advantage of level-wise tree is
that it can be designed into parallel learning easily. However,
given the same stopping criteria (e.g., number of leaves), best-
first tree is expected to achieve better loss than level-wise tree.
The major drawback of best-first tree, however, is prone to
overfitting. Thus, LightGBM controls the maximum tree
depth.
Besides the best-first tree, LightGBM makes some
technical modifications that speed up the training and save
computing sources. For example, a histogram-based algorithm
is employed to reduce the computing cost of searching for
optimal tree structure. Moreover, LightGBM makes
considerable improvement in parallel learning to process large
amount of data. Specifically, feature parallel, data parallel and
GPU support are utilized to completing best-first trees
efficiently. Moreover, LightGBM employs advanced
collective communication algorithms which optimize the
network communication issue in parallel learning.
Insert Fig. 1 here
C. OD-LIGHTGBM FOR REJECT INFERENCE
In this subsection, we introduce the proposed reject inference
framework that combines outlier detection and LightGBM.
This approach consists the following three steps.
1) DATA PREPROCESSING
Let 𝐴 and 𝑅 denote the index of accepted and rejected
(4)
applications, respectively. The basic idea of reject inference is
to infer the potential status of rejected loans (𝑦 ) and then
VOLUME XX, 2019
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2019.2927602, IEEE
Access
build a credit scoring model combining the information on
both accepted (𝐱 , y ) and rejected applicants (𝐱 , 𝑦 ). Thus,
we first preprocess the data and separate the data into the
accepted and rejected sets.
Moreover, the features of accepted and rejected data maybe
inconsistent and requires handling missing value issue. There
are several measures to address the problem, among which the
most commonly-used ones are deletion, imputation, and
dummy variables [49]. Deletion and dummy variables
approach are considered unsuitable since list-wise deletion or
dummy variable may hinder the process of reject inference and
column-wise deletion harms the predictability of credit
scoring models. Imputation, therefore, seems to be a rational
choice. Mean imputation and median imputation are two
representative imputation methods. However, we consider it is
not feasible to employ these two methods since some missing
features (i.e., term, interest rate, debt-to-income ratio and total
amount of credit lines) in accepted and rejected loans may be
quite different. Replacing the missing data in rejected
applicants with mean or median of those accepted applicants
may lead to confusing classification boundary and thus, poor
results. Following the work of Xia, et al. [27], we replace the
missing values in rejected samples with 0 to maintain some
dissimilarity between accepted and rejected samples.
Though GBDT-based techniques are seldom affected by
scalar of data, the performances of outlier detection algorithm
used in next subsection may be highly dependent on data
normalization Thus, the features of all data are normalized into
[0,1] after imputation.
2) INFERRING REJECTED LOANS VIA OUTLIER
DETECTION
The core issue of this subsection is to infer rejected loans using
available information. Some techniques have been applied to
handle this task. However, several drawbacks exist for these
popular approaches. First, all these methods cannot modify the
good/bad (GB) ratio of the rejects group flexibly despite the
fact that the efficiency of reject inference is highly dependent
on the accurate estimation of GB ratio for all the applicants
[19]. Second, some approaches hold key assumptions that may
be far away from reality. For example, augmentation assumes
the potential performances of the rejects can be directly
imputed from those of the accepts. However, such an
assumption may be violated due to sample bias. Finally, the
current reject inference techniques are mainly built on
statistical approaches though AI-based method is regarded as
an interesting direction for further studies [16, 27]. Thus, we
aim to employ outlier detection algorithm to overcome these
drawbacks.
Different from those of definitions in prior studies which
apply outlier detection in data preprocessing (e.g., García, et
al. [14] and Tian, et al. [16]), the outlier herein implies a good
applicant who is accidentally rejected due to some incidental
factors. Under this circumstance, the features of outliers are
distinctly different from those of inliers and should be
regarded as good applicants. By these means, the potential
VOLUME XX, 2019
This work is licensed under a Creative Commons Attribution 3.0 License. For more information, see http://creativecommons.org/licenses/by/3.0/.
status of rejects can be determined depending on whether a
sample belongs to outlier or not. We use a simple example in
Fig. 2 to illustrate the mechanism of outlier detection in reject
inference. The triangles and circles denote good and bad
applicants, respectively. The x points imply rejected samples.
The full plane in black herein is the optimal classification
boundary considering only the accepts. In such a case, a
classifier provides sub-optimal performances since a few
samples are misclassified. We further apply outlier detection
algorithm in rejected applicants. Specifically, the outliers are
denoted as good applicants and printed in blue. On contrast,
the inliers are denoted as bad applicants and printed in green.
In this case, an optimal classification boundary can be depicted
in the figure (dashed red plane). The misclassification rate is
reduced after employing outlier detection in reject inference.
Moreover, outlier detection is typically an unsupervised
learning and independent of existing credit scoring models.
Consequently, such a method relaxes the constraints in current
reject inference approaches.
Insert Fig. 2 here
In this paper, we use isolation forest [17] to infer the status
of the rejects. The basic idea of isolation forest, which hugely
different from popular outlier detection algorithm such as local
outlier factor or stochastic outlier selection, is to explicitly
identify outliers instead of profiling normal samples. Isolation
forest is inherently an ensemble tree approach. In the root node
of each binary search tree, a feature and a cut-off value in the
range of the feature is randomly selected. The data points are
therefore split into the child nodes of the tree. The splitting
continue until there is only one data point is one node or the
maximum tree height is reached. The subsequently issue lies
on how to quantify the degree of anomaly for a given data
point. Theoretically, the outliers are much less frequently than
inliers and have different patterns from normal samples. Thus,
the outliers tend to be identified closer to the root node (shorter
path length) than inliers. By these means, the anomaly score 𝑠
of isolation forest is described as follows:
( ( ))
s(x, n) = 2 ( ) , (7)
where ℎ(𝑥) is the path length of sample 𝑥 . A detailed
definition of path length can be found in Liu, et al. [17]. 𝑐(𝑛)
denotes the average path length of unsuccessful search in
binary search tree. 𝐸(ℎ(𝑥)) herein implies the average of
ℎ(𝑥) from an ensemble of isolation trees. Each sample can be
given an anomaly score following Eq. (7) and decisions can
be made based on the score. If a sample has an anomaly score
very closet 1, it is possibly an outlier. A sample with anomaly
score below 0.5 indicates an inlier. Specifically, if the anomaly
scores of all the samples are approximately 0.5, then the entire
sample does not have distinct anomalies. Using the anomaly
scores, we can also adjust the proportion of outliers (or CR)
based on their rankings.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2019.2927602, IEEE
Access
Yufei Xia: A Novel Reject Inference Model Using Outlier Detection and Gradient Boosting Technique in Peer-to-Peer Lending (June
2019)
Given an outlier ratio, we execute the isolation forest
algorithm to detect the outliers in the rejected dataset.
Subsequently, we label the status of outliers as good applicants,
whereas denote the inliers as risky applicants.
3) CLASSIFICATION USING LIGHTGBM
We now have status of all the samples. Then we can use the
whole dataset (including both the accepts and rejects) to train
a LightGBM model. Given a new data point, the LightGBM
model can return the PD and the class that the applicants
should be labeled as. Finally, the model performances are
evaluated in terms of predictability and computational
complexity.
IV. DATA
In this paper, we use two real-world datasets (i.e., Lending
Club dataset and We.com dataset) to perform the experiment.
The Lending Club dataset is a public one, which can be freely
downloaded from the official website of Lending Club. The
We.com dataset is a private one that accumulated via web
crawler algorithm.
Lending Club is the world’s largest online marketplace for
social lending. To maintain the transparency of business,
Lending Club have released the real transaction documents
after eliminating privacy information. In this paper, we extract
the data from January 2009 to December 2012 since even 60-
months loans will have clear status till now. After removing
the samples with obvious error or omit, the whole dataset
contains samples, among which 91,825 are accepted
applicants and 716,505 are rejected. Two critical problems
must be addressed before using Lending Club dataset. First,
the rejected samples contain only five features, namely, loan
amount, FICO score, DTI ratio, region, and employment
length. However, as shown in Serrano-Cinca, et al. [8], some
powerful predictors are not included in the rejected dataset,
which may further hinder the predictability of credit scoring
models. Thus, we add some extra variables, namely revolving
line utilization rate, number of open credit lines, term, and
interest rate. Second, the region information is recorded in text
and requires processing. Following the suggestion of Xia, et
al. [27], we convert the address state into ordinal number
according to the prior default rate.
We.com is a mainstream P2P lending platform in China. The
We.com dataset is a comparatively small dataset, including
1,489 accepted applicants and 2,968 rejected ones after
removing observations with obvious mistakes. One of the
defining characteristics of We.com dataset is that the features
of the accepts and the rejects are consistent. Thus, imputation
is no longer required for this dataset.
The summary statistics and correlation matrixes of the two
datasets are shown in Table 2 and 3, respectively. From the
two tables, we can draw that features of the accepts and the
rejects have very different distributions, which inversely
implies that the missing mechanism is probably MNAR.
This work is licensed under a Creative Commons Attribution 3.0 License. For more information, see http://creativecommons.org/licenses/by/3.0/.
Insert Tables 2 and 3 here
V. EXPERIMENT SETUP AND RESULTS
A. EXPERIMENT DESIGN
In this subsection, we aim to compare the proposed reject
inference model with benchmark models in both fields of
credit scoring and reject inference, such as LR, RF, SVM, and
S3VM. Furthermore, we conduct a sensitivity analysis on the
proposed model under different settings of rejection rate and
contamination rate. Besides, we compare the computational
costs of the proposal and benchmark models. As suggested by
Li, et al. [13], we carry out the experiment as follows:
Step 1: Sample accepted and rejected samples randomly,
whose number is denoted as N and N , respectively. To
mitigate the effect of sample bias, we use a different random
seed for each sample.
Step 2: Split the accepted samples into training set and test
set, the proportion of which 701% and 30%.
Step 3: Building supervised credit scoring models, namely
LR, RF, and SVM, using the training set of the accepted
dataset, and S3VM, OD-LightGBM using both the training set
of the accepted dataset and the rejected dataset with potential
labels.
Step 4: Utilizing the classifiers built up in Step 3 to predict
the PD and label of samples in test set. Then compare the
model performances of different approaches.
To make fair comparison, we repeat the experiment for 50
times and the results are computed as the average value of
these experiments. Besides, we set N = 2000 following the
work of Xia, et al. [27]. To test the influence of rejection rate
on model performances, we set N as different values, ranging
from 2000 to 20000. The whole experiment is coded in Python
2.7 and the scikit-learn package [50], and LightGBM package
[18] are employed.
B. EVALUATION METRICS
Several evaluation metrics have been employed to measure
model performance. Accuracy, which is the proportion of
correctly classified samples among the total number of
samples, is used to evaluate the label predictability of models.
Following the work of Bequé and Lessmann [51], we adjust
the cut-off value of accuracy according to the GB ratio in
training set. Moreover, Area under Receiver Operation
Characteristic Curve (AUC) and H measure are utilized to
measure the discriminative capability of models. Specifically,
H measure overcomes some drawbacks of traditional AUC
measure and is applied in several recent studies. Finally, we
consider computational complexity since it is critical for large-
size problem. Computational time consumption is employed
as an indicator of computation complexity. To ensure the
comparability of the running time, all the experiments are
executed in the same desktop PC with AMD Rayzen 1800X
VOLUME XX, 2019
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2019.2927602, IEEE
Access
8-core 3.6GHz CPU, 24 GB RAM, Nvidia GeForce GTX
1700 GPU, and 64-bit Windows 10 operating system.
C. HYPER-PARAMETER OPTIMIZATION
AI-based classifiers, especially SVM and LightGBM used in
our experiment, usually have several hyper-parameters that
largely influence model performance. Thus, we need to tune
hyper-parameters of these models. For SVM and S3VM, since
we employ a radial basis function kernel, the hyper-parameters
and their functions are as follows:
Soft margin hyper-parameter 𝐶 : controls the trade-off
between training error and margin maximization.
Kernel hyper-parameter 𝛾: determines the flexibility of the
decision boundary.
The hyper-parameters of LightGBM are very similar with
those of GBDT. Moreover, LightGBM provides some more
hyper-parameters that control overfitting and lead to
promising performances. The detailed hyper-parameters of
LightGBM are summarized as follows:
Learning rate: shrinks the contribution of each base model.
Number of iterations: number of boosting iterations.
Maximum leaves: limit the maximum child leaves of single
base model.
Bagging fraction: the proportion of training data that sued
to build up base model.
Maximum number of bins: limit the maximum number of
bins that feature values will be bucketed in.
Overall, all these hyper-parameters are tuned via a state-of-
the-art Bayesian hyper-parameter optimization algorithm [52].
Such an approach is empirically more effective than popular
grid search. A 5-fold cross-validation AUC value is employed
as the evaluation metric to select the optimal hyper-parameter
setting of the aforementioned models.
D. Results
1) ANALYSIS OF PREDICTIVE PERFORMANCE
The numerical results of the proposal and benchmark models
are shown in Tables 4 and 5 for Lending Club and We.com
dataset, respectively. Regarding the classification accuracy,
we provide extra information on type I and type II error rates.
The two types of error rates reflect the predictability for good
and bad applicants, respectively. Note that type II error
(misclassifying a real bad applicant) is more expensive than
type II error (treat a good applicant as risky one) [51].
Moreover, we set N = N = 2000 for Lending Club dataset.
We maintain the original N and N for We.com dataset due
to its relatively small size. For both datasets, the CR is
determined as 0.2. The best-performing model for each
evaluation metric is highlighted in bold, and the second-best
model is denoted in italics. We have also employed a
parametric statistical hypothesis test (i.e., paired t test) to
perform between the best-performing and second-best models.
Though Lessmann, et al. [11] claimed that the assumptions of
parametric tests are typically violated in multiple comparisons
of credit scoring models, we execute extra tests to show that
one of the strictest assumptions, namely the distribution of the
VOLUME XX, 2019
This work is licensed under a Creative Commons Attribution 3.0 License. For more information, see http://creativecommons.org/licenses/by/3.0/.
differences between the two model performances should be
approximately normally distributed, is held for our cases. For
each evaluation metric, we report the Shapiro-Wilk test
statistic for the difference between best model and second-best
model.
We first focus on the Lending Club dataset. It can be drawn
from Table 4 that the proposed OD-LightGBM provided a
promising result. Concretely, it outperformed all other models
regarding AUC and H measure over the four datasets. The
second-best model is LightGBM for AUC and H measure.
OD-LightGBM also achieved the best accuracy, type I and
type II error rates in 2009 dataset. The results of normality test
showed that for every evaluation measure, the difference
between best and second-best model followed normal
distribution, indicating that it is feasible to employ a paired-t
test for the two models. The results of parametric tests also
implied that OD-LightGBM significantly outperformed the
benchmarks for AUC and H measure in 2009, 2011, and 2012
dataset. Though our proposal became the second-best model
for accuracy-related measures, the gap is negligible based on
the paired-t tests.
Insert Tables 4 and 5 here
We then focus on We.com dataset. From Table 5 we can
conclude that the gap between different models are
comparatively minor for this dataset. In such as dataset, the
proposed OD-LightGBM performed significantly better than
the second-best model (i.e., supervised LightGBM) regarding
AUC and H measure. In terms of accuracy and type I error rate,
OD-LightGBM performed slightly better than supervised
LightGBM. However, the paired-t test was not performed due
to the corresponding difference did not pass normality test. For
type II error rate, supervised LightGBM outperforms OD-
LightGBM a bit whereas the non-normal distribution of
differences between the two models hinders the further
significance test.
To summarize, the proposed OD-LightGBM perform the
best among the benchmark models on average. Concretely, it
achieved surprisingly good results in terms of AUC and H
measure, whereas it became the second-best model for
accuracy, type I and type II error rates in most cases. It is worth
mentioning that the accuracy, type I and type II error rates will
change if the threshold described in subsection B of Section V
is modified. The sound results on AUC imply that the
performances of OD-LightGBM on these measures can be
further improved by modifying the threshold.
2) ANALYSIS OF COMPUTATIONAL COMPLEXITY
Apart from predictive ability, we analyze computational
complexity of models in this subsection. Computational cost
indicates the computational resources consumed in building
credit scoring models. Computational complexity is drawing
much attention in practical credit scoring, especially for large-
sized problems. For example, P2P lending platforms are
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2019.2927602, IEEE
Access
Yufei Xia: A Novel Reject Inference Model Using Outlier Detection and Gradient Boosting Technique in Peer-to-Peer Lending (June
2019)
expected to provide the credit score of a new applicant with
low latency. Similarly, low computational complexity means
the credit scores can be updated frequently. Thus, a perfect
credit scoring model should achieve sound predictive
performances and meanwhile maintain relative efficiency.
Since reject inference model has an additional step, namely
inferring the potential status of the rejects, the computational
complexity of reject inference model may increase. However,
the application of multi-thread computing and GPU
computing help to tackle the complexity issue. Specifically,
the tree-based ensemble algorithms, such as random forest and
isolation forest, can be quickly executed using multiple
threads of CPU. Moreover, LightGBM support GPU
computing, which has limitless potential in handling large-
sized data. To the best of our knowledge, no prior studies have
applied GPU acceleration in building up GBDT-based credit
scoring models.
We consider five model (i.e., LR, RF, S3VM, OD-
LightGBM (CPU version), and OD-LightGBM (GPU version)
regarding computational complexity. The results of average
computational time (in seconds) of different approaches are
summarized in Table 6. Recall that in the original setting,
N = 2000 . Table 6 clearly shows that LR requires least
computation costs in training and predicting, whereas its
predictive performance is unsatisfactory as indicated in Tables
4 and 5. RF is slightly slower than LR but achieves much
better prediction accuracy than LR. Such a result, again,
supports the claim of Lessmann, et al. [11], who advocated
replacing the industry benchmark LR with RF. Despite OD-
LightGBM consumes more time than LR and RF, the running
time can be dramatically save using GPU computing,
decreasing by nearly 50%.
Insert Table 6 here
Though S3VM is reported as an effective tool for reject
inference [13], SVM, as a kernel method, is easily affected by
curse of dimension and is not suitable to large-scale
applications. To confirm such a claim, we test the efficiency
of reject inference models under different settings of N . The
results of running time are summarized in Table 6. Note that
the maximum running time is determined as 10,000s. From the
results we can conclude that the proposed OD-LightGBM not
only provides the best AUC, but also consumes limited
computation resources. It is worth mentioning that S3VM is
unable to handle large-sized data within a reasonable time,
whereas OD-LightGBM can handle them quietly efficiently.
In summary, compared with the benchmark LR and RF
methods, our proposal provides better performances at the cost
of extra computational resources, whereas the application of
GPU-computing can partially remedy such a drawback. When
comparing the state-of-the-art S3VM with the proposed OD-
LightGBM model, our approach is more accurate and easy-to-
This work is licensed under a Creative Commons Attribution 3.0 License. For more information, see http://creativecommons.org/licenses/by/3.0/.
implement. Thus, the proposed approach has great potential in
some large-sized reject inference applications.
3) SENSITIVITY ANALYSIS
In previous subsections, the experiments are conducted under
some parameter settings. For example, N and N is
determined beforehand. The contamination rate is also set as
0.2. Furthermore, the proposed “outlier detection + classifier”
reject inference framework is limited to a specific combination
(i.e., isolation forest + LightGBM). It is necessary to perform
a sensitivity analysis for the following three reasons: first, the
accepted data is very precious, whereas the rejected data is
comparatively valueless since most of applicants are rejected
in P2P lending domain. Thus, reject inference is especially
attractive since it provides an approach to convert the useless
data into knowledge. It is interesting to explore the optimal N
since it will benefit practical reject inference in P2P lending.
Second, CR potentially indicates the good/bad proportion of
the rejected dataset. The change of CR, therefore, will affect
the performances of reject inference model theoretically.
Consequently, it is important to investigate in the mechanism
of CR on model performances and at best, provide a rule-of-
thumb to guide practical implementation. Finally, the
efficiency of the proposed reject inference framework may be
questioned if only applying one combination of outlier
detection algorithm and classifier. To check the robustness of
our proposal, we integrate local outlier factor (LOF) and
Histogram-based Outlier Detection (HBOS) with RF and
LightGBM.
Insert Table 7 here
To reach a reliable conclusion, we first examine the effect
of N on model performances. Table 7 reports the
performances of OD-LightGBM under various combinations
of N and CR. The results clearly show that the performances
of OD-LightGBM is quite stable, outperforming the
benchmarks under various parameter settings. For each dataset,
the average AUC and H measure are calculated for different
N . From these average values, we can draw two main
conclusions. First, the differences between models under
various N are minor. Specifically, the gap between the best
and the worst model does not exceed 0.001, which implies that
N can hardly affect model performances. Second, despite the
negligible difference, an inverse U-shape can be observed for
AUC and H-measure of each dataset. In other words, the
model performances increase until a certain N is reached.
Though varying from datasets, a candidate N ranging from
4000 to 8000 is encouraged since it achieves a promising
tradeoff between model performances and computational
costs. Though N = 20000 is quite close to the real rejection
rate in Lending Club, our proposal achieved sub-optimal
performances under such a parameter setting. This implies that
a sampling of the rejected data benefits both model
performances and computational costs.
VOLUME XX, 2019
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2019.2927602, IEEE
Access
We then analyze the influences of CR. As shown in Table
7, the optimal CR varies for different datasets. Concretely, the
differences of model performances are minor across CRs for
Lending Club 2009 dataset and CR = 0.3 achieves the best
results. However, for the Lending Club 2010, 2011, and 2012
dataset, the optimal CR equals to 0.05. Such a result indicates
that the optimal choice of CR should be approximately 0.05 in
most cases. The low optimal CR also demonstrates that an
overwhelming majority of the rejects are potentially risky
applicants.
Insert Table 8 here
The results of LOF-RF, LOF-LightGBM, HBOS-RF, and
HBOS-XGBoost are summarized in Table 8. We herein use
the original parameter setting for these reject inference
methods, namely N = N = 2000. From the table we can
conclude that the proposed reject inference framework
provided robust performances under different combinations of
outlier detection algorithm and classifiers. Specifically, the
application of LOF and HBOS improved the AUC and H
measure for all the datasets relative to the supervised RF or
LightGBM. Our proposal also increased accuracy and
decreased type I and II error rates for Lending Club 2011, 2012,
and We.com dataset. These results imply that the proposed
“outlier detection + classifier” reject inference framework can
be potentially further improved by employing powerful outlier
detection algorithms or classification techniques.
VI. CONCLUSIONS AND FUTURE RESEARCH
P2P lending is growing rapidly across the world and the
information asymmetry is a core issue in this newly
established industry. Credit scoring, as an efficient solution to
information problems, is drawing much attention for both
academia and industry. The credit scoring models are typically
built with the accepted applicants, which may cause sample
bias and hinder model performances. Recently, reject
inference methods that employ the rejected data to mitigate
sample bias have become research hotspot in credit scoring.
This paper develops a novel reject inference framework based
on outlier detection technique and LightGBM (i.e., OD-
LightGBM). Our proposal has the following four potential
advantages. First, it is a comparatively flexible framework
since the outlier detection algorithm and the classifiers can be
freely changed. Second, unlike popular S3VM technique, a
state-of-the-art classification algorithm (i.e., LightGBM) is
employed to discriminate the good and risky applicants. Third,
the proposed model can alleviate the dramatic computational
costs of commonly-used S3VM. Finally, the GPU support of
LightGBM guarantees the high efficiency of handling large-
sized problems.
The proposed model is validated on two real-world credit
datasets of P2P lending. The numeric results of predictive
performances strongly demonstrate the superiority of our
proposal in terms of discriminative capability. The analysis of
computational cost shows the feasibility of our proposal in
VOLUME XX, 2019
This work is licensed under a Creative Commons Attribution 3.0 License. For more information, see http://creativecommons.org/licenses/by/3.0/.
handling large-sized problems using GPU computing relative
to classic reject inference methods. Thus, OD-LightGBM
shows great potentials as a powerful reject inference method
in practice. The sensitivity analysis further illustrates that the
proposed reject inference framework can achieve robust
performances under different parameter settings and maybe
improved by employing advanced outlier detection algorithms
and classifiers.
For future studies, one can explore the following three
directions. First, though a random sampling of the rejects
improves the performances of reject inference model is
demonstrated in sensitivity analysis, one may explore whether
the model would be further optimized using some special
sampling techniques. Second, due to the flexibility of our
proposal, powerful outlier detection algorithms, as well as
classification techniques can be incorporated in such a reject
inference framework. Finally, it is interesting to discuss the
efficiency of reject inference from a cost-benefit perspective.
REFERENCES
[1] L. Chen, "From fintech to finlife: the case of fintech
development in China," China Economic Journal, vol. 9, no. 3,
pp. 225-239, 2016.
[2] R. Emekter, Y. Tu, B. Jirasakuldech, and M. Lu, "Evaluating
credit risk and loan performance in online Peer-to-Peer (P2P)
lending," Applied Economics, vol. 47, no. 1, pp. 54-70, 2015.
[3] A. Mild, M. Waitz, and J. Wöckl, "How low can you go?—
Overcoming the inability of lenders to set proper interest rates
on unsecured peer-to-peer lending markets," Journal of Business
Research, vol. 68, no. 6, pp. 1291-1305, 2015.
[4] S. Freedman and G. Z. Jin, "The information value of online
social networks: lessons from peer-to-peer lending,"
International Journal of Industrial Organization, vol. 51, pp.
185-222, 2017.
[5] J. E. Stiglitz and A. Weiss, "Credit rationing in markets with
imperfect information," The American economic review, vol. 71,
no. 3, pp. 393-410, 1981.
[6] C. R. Everett, "Group Membership, Relationship Banking and
Loan Default Risk: The Case of Online Social Lending,"
Banking and Finance Review, vol. 7, no. 2, pp. 15-54, 2015.
[7] M. Malekipirbazari and V. Aksakalli, "Risk assessment in social
lending via random forests," Expert Systems with Applications,
vol. 42, no. 10, pp. 4621-4631, 2015.
[8] C. Serrano-Cinca, B. Gutierrez-Nieto, and L. López-Palacios,
"Determinants of default in P2P lending," PloS one, vol. 10, no.
10, p. e0139427, 2015.
[9] Y. Guo, W. Zhou, C. Luo, C. Liu, and H. Xiong, "Instance-Based
Credit Risk Assessment for Investment Decisions in P2P
Lending," European Journal of Operational Research, vol. 249,
no. 2, pp. 417-426, 2016.
[10] T. Verbraken, C. Bravo, R. Weber, and B. Baesens,
"Development and application of consumer credit scoring
models using profit-based classification measures," European
Journal of Operational Research, vol. 238, no. 2, pp. 505-513,
2014.
[11] S. Lessmann, B. Baesens, H.-V. Seow, and L. C. Thomas,
"Benchmarking state-of-the-art classification algorithms for
credit scoring: An update of research," European Journal of
Operational Research, vol. 247, no. 1, pp. 124-136, 2015.
[12] J. M. Liberti and M. A. Petersen, "Information: Hard and soft,"
Review of Corporate Finance Studies, vol. 8, no. 1, pp. 1-41,
2018.
[13] Z. Li, Y. Tian, K. Li, F. Zhou, and W. Yang, "Reject inference
in credit scoring using Semi-supervised Support Vector
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2019.2927602, IEEE
Access
Yufei Xia: A Novel Reject Inference Model Using Outlier Detection and Gradient Boosting Technique in Peer-to-Peer Lending (June
2019)
Machines," Expert Systems with Applications, vol. 74, pp. 105-
114, 2017.
[14] V. García, A. Marqués, and J. S. Sánchez, "On the use of data
filtering techniques for credit risk prediction with instance-based
models," Expert Systems with Applications, vol. 39, no. 18, pp.
13267-13276, 2012.
[15] T. Harris, "Credit scoring using the clustered support vector
machine," Expert Systems with Applications, vol. 42, no. 2, pp.
741-750, 2015.
[16] Y. Tian, Z. Yong, and J. Luo, "A new approach for reject
inference in credit scoring using kernel-free fuzzy quadratic
surface support vector machines," Applied Soft Computing, vol.
73, pp. 96-105, 2018.
[17] F. T. Liu, K. M. Ting, and Z.-H. Zhou, "Isolation forest," in 2008
Eighth IEEE International Conference on Data Mining, 2008,
pp. 413-422: IEEE.
[18] G. Ke et al., "LightGBM: A Highly Efficient Gradient Boosting
Decision Tree," in Advances in Neural Information Processing
Systems, 2017, pp. 3149-3157.
[19] J. Crook and J. Banasik, "Does reject inference really improve
the performance of application scoring models?," Journal of
Banking & Finance, vol. 28, no. 4, pp. 857-874, 2004.
[20] J. Banasik and J. Crook, "Reject inference, augmentation, and
sample selection," European Journal of Operational Research,
vol. 183, no. 3, pp. 1582-1594, 2007.
[21] A. Feelders, "Credit scoring and reject inference with mixture
models," International Journal of Intelligent Systems in
Accounting, Finance & Management, vol. 9, no. 1, pp. 1-8, 2000.
[22] J. Banasik and J. Crook, "Reject inference in survival analysis
by augmentation," Journal of the Operational Research Society,
vol. 61, no. 3, pp. 473-485, 2010.
[23] S. Maldonado and G. Paredes, "A Semi-supervised Approach for
Reject Inference in Credit Scoring Using SVMs," in Industrial
Conference on Data Mining, 2010, pp. 558-571: Springer.
[24] G. G. Chen and T. Åstebro, "Bound and collapse Bayesian reject
inference for credit scoring," Journal of the Operational
Research Society, vol. 63, no. 10, pp. 1374-1387, 2012.
[25] Y. Xia, C. Liu, and N. Liu, "Cost-sensitive boosted tree for loan
evaluation in peer-to-peer lending," Electronic Commerce
Research and Applications, vol. 24, no. July–August pp. 30–49,
2017.
[26] A. Kim and S.-B. Cho, "An ensemble semi-supervised learning
method for predicting defaults in social lending," Engineering
Applications of Artificial Intelligence, vol. 81, pp. 193-199, 2019.
[27] Y. Xia, X. Yang, and Y. Zhang, "A rejection inference technique
based on contrastive pessimistic likelihood estimation for P2P
lending," Electronic Commerce Research and Applications, vol.
30, pp. 111-124, 7// 2018.
[28] D. M. Hawkins, Identification of outliers. Springer, 1980.
[29] V. Hodge and J. Austin, "A survey of outlier detection
methodologies," Artificial intelligence review, vol. 22, no. 2, pp.
85-126, 2004.
[30] A. Srivastava, A. Kundu, S. Sural, and A. Majumdar, "Credit
card fraud detection using hidden Markov model," IEEE
Transactions on dependable and secure computing, vol. 5, no. 1,
pp. 37-48, 2008.
[31] S. Panigrahi, A. Kundu, S. Sural, and A. K. Majumdar, "Credit
card fraud detection: A fusion approach using Dempster–Shafer
theory and Bayesian learning," Information Fusion, vol. 10, no.
4, pp. 354-363, 2009.
[32] Z. Zhang, G. Gao, and Y. Shi, "Credit risk evaluation using
multi-criteria optimization classifier with kernel, fuzzification
and penalty factors," European Journal of Operational Research,
vol. 237, no. 1, pp. 335-348, 2014.
[33] H. Liu, S. Shah, and W. Jiang, "On-line outlier detection and data
cleaning," Computers & chemical engineering, vol. 28, no. 9, pp.
1635-1647, 2004.
[34] S. Kotsiantis, D. Kanellopoulos, and P. Pintelas, "Data
preprocessing for supervised leaning," International Journal of
Computer Science, vol. 1, no. 2, pp. 111-117, 2006.
[35] R. Setiono, A. Azcarraga, and Y. Hayashi, "Using sample
selection to improve accuracy and simplicity of rules extracted
This work is licensed under a Creative Commons Attribution 3.0 License. For more information, see http://creativecommons.org/licenses/by/3.0/.
from neural networks for credit scoring applications,"
International Journal of Computational Intelligence and
Applications, vol. 14, no. 04, p. 1550021, 2015.
[36] 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.
[37] M. Wang, X. Zheng, M. Zhu, and Z. Hu, "P2P lending platforms
bankruptcy prediction using fuzzy SVM with region
information," in 2016 IEEE 13th International Conference on e-
Business Engineering (ICEBE), 2016, pp. 115-122: IEEE.
[38] A. Byanjankar, M. Heikkilä, and J. Mezei, "Predicting credit risk
in peer-to-peer lending: A neural network approach," in
Computational Intelligence, 2015 IEEE Symposium Series on,
2015, pp. 719-725: IEEE.
[39] F. Tan, X. Hou, J. Zhang, Z. Wei, and Z. Yan, "A deep learning
approach to competing risks representation in peer-to-peer
lending," IEEE transactions on neural networks and learning
systems, 2018.
[40] 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.
[41] X. Ye, L.-a. Dong, and D. Ma, "Loan Evaluation in P2P Lending
based on Random Forest Optimized by Genetic Algorithm with
Profit Score," Electronic Commerce Research and Applications,
2018.
[42] W. Li, S. Ding, Y. Chen, and S. Yang, "Heterogeneous ensemble
for default prediction of peer-to-peer lending in china," IEEE
Access, vol. 6, pp. 54396-54406, 2018.
[43] 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," Electronic Commerce
Research and Applications, vol. 31, pp. 24-39, 2018.
[44] Y. Xia, C. Liu, B. Da, and F. Xie, "A novel heterogeneous
ensemble credit scoring model based on bstacking approach,"
Expert Systems with Applications, vol. 93, pp. 182-199, 2018.
[45] J. H. Friedman, "Greedy Function Approximation: A Gradient
Boosting Machine," Annals of Statistics, vol. 29, pp. 1189-1232,
2000.
[46] C. M. Bishop, Pattern recognition and machine learning.
springer, 2006.
[47] D. Wang, Y. Zhang, and Y. Zhao, "LightGBM: an effective
miRNA classification method in breast cancer patients," in
Proceedings of the 2017 International Conference on
Computational Biology and Bioinformatics, 2017, pp. 7-11:
ACM.
[48] H. Shi, "Best-first decision tree learning," The University of
Waikato, 2007.
[49] R. Anderson, The credit scoring toolkit: theory and practice for
retail credit risk management and decision automation. Oxford
University Press, 2007.
[50] F. Pedregosa et al., "Scikit-learn: Machine learning in Python,"
Journal of machine learning research, vol. 12, no. Oct, pp. 2825-
2830, 2011.
[51] A. Bequé and S. Lessmann, "Extreme Learning Machines for
Credit Scoring: An Empirical Evaluation," Expert Systems with
Applications, vol. 86, pp. 42-53, 2017.
[52] J. S. Bergstra, R. Bardenet, Y. Bengio, and B. Kégl, "Algorithms
for hyper-parameter optimization," in Advances in Neural
Information Processing Systems, 2011, pp. 2546-2554.
VOLUME XX, 2019