https://doi.org/10.31449/inf.v46i7.4247 Informatica 46 (2022) 145 –152 145 
Risk Prediction of Enterprise Credit Financing Using Machine 
Learning 
Jingjing Gu 
Sichuan College of Architectural Technology, Deyang, Sichuan 618000, China 
Email: gu41915074@163.com 
Keywords: Machine learning; Enterprise financing; Credit risk; Countermeasure neural network; Least squares 
support vector machine 
Received: June 18, 2022 
For the credit financing risk of small, medium-sized and micro enterprises, a risk prediction model of 
enterprise credit financing management based on antagonistic neural network and least squares support 
vector machine is proposed. The data samples are processed combined with antagonistic neural network, 
and the risk prediction of enterprise credit financing is realized by using least squares support vector 
machine model. Compared with many classification models, the effectiveness of the least squares 
support vector machine model is verified. The results show that the accuracy rate of the least squares 
support vector machine model is 90.15%, the recall rate is 85.63%, and the prediction error rates of the 
model in default and non-default are 6.48% and 3.09% respectively, which is better than BP neural 
network, random forest algorithm and Gaussian naive Bayesian algorithm. The least squares support 
vector machine model can effectively and accurately predict the enterprise credit financing risk, provide 
scientific and efficient risk early warning for the enterprise credit financing control, and provide 
technical support for preventing and reducing the occurrence of loan default. 
Povzetek: Narejena je metoda strojnega učenja za ugotavljanje tveganj za mala in mikro podjetja. 
 
1 Introduction 
At present, China is in the social background of 
economic transformation and upgrading. The country has 
put forward the development strategy of innovation and 
entrepreneurship for all, encouraged and supported 
entrepreneurs to actively explore the market, boosted the 
development speed of the national economy and 
consumption, and solved important livelihood issues 
such as employment [1]. With the emergence of startups, 
how to promote the sound development of small, 
medium-sized and micro enterprises has become a key 
issue in the national economic development. It is 
required to increase support for enterprises and help 
enterprises solve important financing problems in the 
early development of entrepreneurship [2-3]. The 
traditional enterprise bank loans have complex 
procedures, long review process, and high borrowing 
threshold, which are difficult to meet the borrowing 
needs of enterprises. The development of Internet finance 
has proposed new solutions to the financing problems of 
small, medium-sized and micro enterprises. The 
operation process of Internet financial lending is simple, 
and it has the advantages of low entry threshold and low 
difficulty in lending. It greatly makes up for the lack of 
micro lending for enterprises in the traditional financial 
industry, provides new financial channels for enterprise 
financing, and solves the problems of small, medium-
sized and micro enterprises with difficult and slow loans  
 
[4-5]. However, the entry threshold of Internet finance 
micro lending is low. While increasing the possibility of 
enterprise financing, it also increases the default risk of  
 
enterprise credit financing [6]. In addition, the 
development of the Internet finance micro lending 
platform is relatively short, and there are still problems 
such as lack of management experience and imperfect 
operation structure, which are prone to the situation that 
the loan funds cannot be recovered. It is required to do a 
good job in loan risk management, provide early warning 
for enterprise credit financing risk control while reducing 
the operating risk of the lending platform, and promote 
the common healthy development of the Internet finance 
micro lending platform and small, medium-sized and 
micro enterprises. Therefore, aiming at the problem of 
credit financing risk management of small and medium-
sized micro enterprises, this paper proposes a prediction 
model of enterprise credit financing risk based on 
confrontation neural network and least squares support 
vector machine model, expecting to predict and warn the 
default risk of enterprise credit financing. 
2 Related works 
Facing the problems of unbalanced data samples and 
small amount of data, many scholars use various means 
such as anti-neural network to process data samples. Xu 
y and other scholars proposed a feature data processing 

146 Informatica 46 (2022) 145 –152 J. Gu 
system suitable for neural network calculation for the 
processing of medical feature data, and took the network 
traffic data of medical Internet of things as the object to 
verify the effectiveness of the data processing system in 
convolution neural network and recursive neural network 
[7]. 
H Li and his team build a face anti-deception 
recognition system. When the sample size of training 
data is small, they use the deep learning advantage of 
neural network to extract the characteristic information 
of the original sample, so as to increase the richness of 
attack samples [8]. When solving the problem of medical 
image diagnosis, scholars such as junior F, aiming at the 
imbalance of medical data and few data points, proposed 
a generation anti-neural network pruning algorithm based 
on evolutionary strategy and multi criteria decision-
making, which uses the anti-neural network to obtain 
information features from the original data distribution 
and synthesize more data for problem-solving training 
[9]. Aiming at the problem of synthetic computed 
tomography, scholars such as green a j use the anti-
neural network to process and analyze the virus data, 
convert the three-dimensional structure information of 
the virus into weighted points, and train the neural 
network to realize the effective prediction of the toxicity 
of untested chemicals [10]. 
The least squares support vector machine model can 
effectively deal with nonlinear prediction problems. Tian 
C and other scholars introduced the least squares support 
vector machine model into the ship collision conflict 
prediction problem and proposed a least squares support 
vector machine model based on empirical mode 
decomposition and quantum behavior particle swarm 
optimization, combined with the time series of ship 
collision, the effective prediction of ship collision is 
realized [11]. Zhang X and Ge Z applied the least squares 
support vector machine model to optimize the local 
target set parameters in the field of industrial soft sensing, 
combined with the distribution estimation algorithm to 
reduce the prediction error of the target set, and carried 
out the prediction and estimation experiment of CO2 
content on the cloud computing platform. The results 
show that the improved least squares support vector 
machine model is optimized and feasible [12]. Ge Q and 
his team proposed a combined power load forecasting 
method based on particle swarm optimization algorithm 
and least squares support vector machine for high-
performance industrial power load forecasting, and 
combined with K-means clustering algorithm for data 
classification to realize high-precision prediction of 
industrial power load [13]. Facing the prediction problem 
of boiler NOx emission, Chengang team proposed an 
emission prediction model based on least squares support 
vector machine, which was improved combined with 
whale optimization algorithm to optimize the kernel 
function width and penalty factor of the model. The 
experimental results show that the integrated model has 
high stability and prediction accuracy [14]. Zhao Z and 
other scholars introduced the least squares support vector 
machine model into the ultrafast photovoltaic power 
prediction problem, combined with the single iterative 
gray wolf optimization algorithm for photovoltaic power 
prediction, and optimized the super parameters by 
combining iterative local search and adaptive differential 
evolution to improve the prediction accuracy of the 
model [15]. 
To sum up, the antagonistic neural network can 
effectively deal with the imbalance of data samples, 
enrich the sample data volume, and the least squares 
support vector machine model has significant advantages 
in nonlinear prediction. Therefore, it is studied to build 
an enterprise credit financing risk prediction model based 
on the antagonistic neural network and least squares 
support vector mechanism, it is expected to provide 
technical support for enterprise credit risk control. 
3 Construction of risk prediction 
model for enterprise credit 
financing management 
3.1 Data extraction and structured 
processing 
Extract the enterprise's loan application information data 
from the loan institution database, and obtain the basic 
information of the legal person, loan qualification 
examination data, enterprise operation data and loan 
application product information of the loan enterprise, 
including the credit record of the loan enterprise, 
enterprise capital flow, enterprise operation years and 
other relevant information. The obtained enterprise loan 
information is structured, the keywords in the data 
information are extracted, and the data variables are 
classified and filled in combination with the data 
dictionary. The data structured processing flow is shown 
in Figure 1. 
 
info
Find location
Filling position
Filling content
Generate padding Find keywords
Data dictionary
Complete filling
 
Figure 1: Data structure processing flow 
In the process of model prediction, the original data 
has certain characteristic differences. If the original data 
is directly used for calculation, the size difference of the 
data value will affect the final operation result. Therefore, 
it is necessary to standardize the original data and control 
the variable value within a certain range through the 
normalization of the sample data, so as to make the data 

Risk Prediction of Enterprise Credit Financing Management … Informatica 46 (2022) 145 –152 147 
samples comparable and facilitate the subsequent 
operation and prediction. The maximum and minimum 
method is used to normalize the data. The formula of the 
maximum and minimum method is as follows: 
min
max min
k
k
xx
x
xx



. (1) 
 
In formula (1), 
k
x is sample data, and 
max
x and 
min
x represent the maximum and minimum values in the 
data sequence respectively. There are many variable 
indicators in the sample data, and redundant irrelevant 
information will affect the operation efficiency and 
prediction effect of the model. Therefore, it is necessary 
to reduce the dimension of the variable indicators, 
process the data variables with the principal component 
analysis method, and screen the indicators in 
combination with the variance contribution rate and 
cumulative square difference contribution rate of the 
variable indicators, Eliminate the variable indicators with 
low correlation with enterprise loan risk. 
3.2 Sample processing based on 
Countermeasure neural network 
Structured processing has preliminarily sorted out the 
information data of loan enterprises, but the positive and 
negative proportion of loan data samples is seriously 
unbalanced, which is not conducive to subsequent risk 
prediction. Therefore, the data samples need to be further 
processed. The research uses the antagonistic neural 
network to generate and reconstruct the data samples, 
uses the construction of the neural network model to 
learn the distribution characteristics of the data samples, 
and combines the generated neural network to generate, 
reconstruct and update the original data samples. This 
paper studies the learning and generation of original data 
samples by using the structure network of automatic 
denoising coder. In the coding process, the automatic 
coder first carries out the spatial mapping of information, 
mapping the input vector 
  1,1
d
x 
 into hidden 
  1,1
d
y 
, 
and d is the spatial dimension. The mapping formula is 
shown as follows: 
  y s wx b  . (2)
 
In equation (2), 
  . s is a nonlinear function, tanh 
function and sigmoid function can be used, w is the 
weight and b is the mapping angle. Then map y to 
reconstruction vector z and decode the data information. 
The mapping formula is shown as follows: 
  '' y s w x b  . (3) 
In equation (3), ' w is the weight and ' b is the 
mapping angle. The loss function of vector 
reconstruction is the cross entropy function, which is 
expressed as follows: 
     
1
, log 1 log 1
d
H k k k k
k
L x z x z x x

     


. (4) 
In formula (4), k is the sample number. The 
Min Max  standardization method is used to 
standardize the original data samples. Set the maximum 
and minimum values of attribute A as MaxA and 
MinA respectively, and map the original value x of A 
to ' x and 
  ' 0,1 x  . The standardization processing 
formula is shown as follows: 
'
data MinA
x
MaxA MinA



. (5) 
In equation (5), data is the sample data. The 
standardized original data samples are simulated and 
reconstructed by using the generated neural network, and 
the false samples with similar characteristics are 
generated by using the mapping processing of input layer, 
hidden layer and output layer, combined with the input 
original samples. The judgment neural network is 
constructed to distinguish the real original data samples 
and the generated false samples. The BP neural network 
is used to distinguish the samples. The input information 
is processed through the forward transmission and back 
propagation of the signal. Combined with the adjustment 
of the network weight and threshold, the error between 
the output of the neural network and the expected value 
is reduced. Let the input vector and output vector of the 
neural network be 
 
12
, , , , 1,2, ,
i
x x x x i m  and 
 
12
, , , , 1,2, ,
k
y y y y k m  respectively, and the 
output of hidden layer neurons is 
       
12
, , , , 1,2, ,
l l l l
s
h h h h s p



, of which p is 
the number of hidden layer neurons in layer l , then the 
input of the i neuron in layer l is expressed as follows: 
   
 
 
1
1
p
ll lj
i ij j i
j
l l
ji
net w h b
h f net










. (6) 
In equation (6), 
  l
ij
w is the connection weight of the 
i neuron in layer l and the j neuron in layer 1 l  , 
  l
i
b 
is the bias of the i neuron in layer l , and 
  . f is the 
activation function of neurons, which is usually a 
nonlinear activation function. The false samples 
generated by the model are marked as 0 and the true 
samples are marked as 1 to form the sample training set 
of BP neural network to judge and distinguish the 
generated samples from the original samples. 
The counter neural network is used to deal with the 
imbalance of the original data samples. Firstly, the 
generated neural network is used to simulate and 
reconstruct the original data samples to generate false 
samples that can confuse the false with the true. 
Combined with the training optimization of the judgment 
neural network, the true and false samples are identified 
and judged, and then the connection weight of the model 
is updated to update the generation and judgment series 
model, The schematic diagram of model update is shown 
in Figure 2. 
 
 

148 Informatica 46 (2022) 145 –152 J. Gu 
Generative 
neural network
Judgment neural 
network
loss function
Back propagation
Generative neural network 
after last training
Judgment neural network 
after last training
False sample 
label inversion
Output Connect Input
 
Figure 2: Schematic diagram of model update. 
While updating the generated neural network, the 
training of the judgment neural network is also updated, 
and the new false samples generated by the updated 
generated neural network are trained again to form the 
confrontation competition under the confrontation neural 
network. In order to avoid the homogenization of 
samples after competition due to the randomness of the 
generation of the original sample matrix, it is studied that 
after a random sample competition, only one sample is 
randomly selected as the generated sample, and the 
sample set is formed through multiple generation and 
extraction. 
3.3 Enterprise credit risk prediction model 
based on LSSVM 
Support vector machine (SVM) is a commonly used 
supervised learning algorithm. It analyzes the potential 
laws of data samples on the basis of known data to 
realize data prediction and pattern recognition. Support 
vector machine model has advantages in small sample 
data and nonlinear problems. Therefore, the research 
introduces the support vector machine model into the 
enterprise credit risk problem, and constructs the 
enterprise credit financing risk prediction model based on 
the least squares support vector mechanism to predict 
and analyze the enterprise credit financing risk. Least 
square support vector machine (LSSVM) inherits the 
structural risk minimization strategy and the basic idea of 
kernel function of traditional SVM model, simplifies the 
calculation process of quadratic programming problem, 
transforms it into the solution of linear equations, and 
improves the operation efficiency of the model. The 
support vector machine model is based on the structural 
risk minimization strategy. Its optimization objective 
function is composed of empirical risk and model 
complexity parameters. Select the function subset in the 
function set structure of the model to minimize the 
confidence interval of the function, and further select the 
function that can minimize the empirical risk in the 
function subset, so as to minimize the upper bound of the 
actual risk. The support vector machine model takes the 
limited sample data information as the starting point to 
find the balance between the model learning ability and 
the complexity of the model, so as to improve the 
generalization ability and comprehensive performance of 
the model. 
LSSVM model adopts the least square method, 
constructs the square loss function of the model through 
the selection of the optimal hyperplane and the square of 
the error, and converts the inequality constraint of SVM 
into linear equation to realize the linear solution of the 
problem. Let the data sample set be 
       
11
, , , , , , ,
i i n n
D x y x y x y  , n represent the 
number of samples, the i input sample is 
i
x , and 
i
y 
represent the label of the input sample. The relaxation 
variable 0
i
  is introduced into the model, and the 
constraints of the model are expressed as follows: 
 
1 , 1,2, ,
T
i i i
y x e i n      . (7) 
In equation (7),  is the normal vector of the 
hyperplane, T is the transpose, and e is the intercept 
term. The constrained optimization problem of the model 
is expressed as follows: 
 
2
,
1
1
2
. . 1 , 0, 1,2, ,
n
b
i
T
i i i
min C
s t y x e i n


  






    


. (8) 
In equation (8), C is the penalty parameter, and the 
values of 0 C  and C are positively correlated with 
the penalty of misclassification. The constrained 
optimization problem is transformed by Lagrange 
method, and the model constrained optimization problem 
is expressed as follows: 
1 1 1
1
1
2
. . 0,0 , 1,2, ,
n n n
T
i i j i j i j
i i j
n
i i i
i
min y y x x
s t y C i n

  

  







   


  

. (9) 
Since it is difficult for linear functions to correctly 
distinguish data samples, nonlinear mapping is 
introduced into the model to realize the mapping 
transformation from low-dimensional space information 
to high-dimensional space, making it linearly separable. 
The schematic diagram of nonlinear mapping is shown in 
Figure 3. 
 
y
x
y
x
Mapping 
 
Figure 3: Nonlinear mapping diagram. 
After introducing nonlinear mapping 
  xx   
into the model, the nonlinear mapping is expressed in the 
form of inner product in the model, so it is not necessary 
to obtain the specific expression of 
  x  , but the 
corresponding model decision function is obtained from 
the inner product form of 
  x  . The kernel function is 
used to complete the inner product operation in the low 
dimensional space. The kernel function is expressed as 
follows: 

Risk Prediction of Enterprise Credit Financing Management … Informatica 46 (2022) 145 –152 149 
     
,
T
i j i j
K x x x x   . (10) 
The choice of kernel function is directly related to 
the performance of the whole LSSVM model. Linear 
kernel function, polynomial kernel function, sigmoid 
kernel function and radial basis function kernel function 
are commonly used. Among them, RBF kernel function 
has the advantages of strong locality and less parameters, 
and has good applicability in large sample data and small 
sample data. Therefore, RBF kernel function is selected 
as the kernel function of the model. The radial basis 
kernel function formula is expressed as follows: 
 
2
2
, exp
2
i
i
xx
K x x



 


. (11) 
In equation (11),  is the kernel function parameter. 
Then the constrained optimization problem of the model 
is expressed as follows: 
 
1 1 1
1
1
,
2
. . 0,0 , 1,2, ,
n n n
i i j i j i
i i j
n
i i i
i
min y y K x x
s t y C i n

  

  







   


  

. (12) 
In equation (12), 
i
 is Lagrange multiplier, 0
i
  . 
The model decision function is expressed as follows: 
       
*
1
sgn * * sgn , *
n
T
i i i
i
f x x e y K x x e   



   




 (13)
 
The operation process of enterprise credit financing 
risk prediction model based on antagonistic neural 
network and least squares support vector machine is 
shown in Figure 4. Firstly, the data samples are 
structured and normalized, and then the generation, 
judgment and update of antagonistic neural network 
solve the problem of unbalanced data samples, Finally, 
the risk prediction model based on least squares support 
vector machine is used to predict and warn the enterprise 
credit financing risk. 
 
Data preprocessing Unbalance treatment Risk prediction
Structured
Normalization
Countermeasure neural network
Generate 
LSSVM
Kernel 
function
 Judge Update
 
Figure 4: Operation process of enterprise credit financing 
risk prediction model. 
4 Verification of risk prediction 
effect 
4.1 Model performance analysis 
The performance of the enterprise credit financing risk 
prediction model constructed by the research institute is 
verified and analyzed. The kernel function parameter  
of the model controls the complexity of the model and 
affects the generalization ability of the model, while the 
regularization parameter  is related to the trade-off 
between the empirical risk and confidence range of the 
model. The parameter settings of the model are shown in 
Table 1. 
 
Duration of 
entrepreneurship 
Freshman Sophomore 
Generative 
neural network 
Number of neurons 
in hidden layer 1 
30 
Number of neurons 
in hidden layer 2 
25 
Judgment neural 
network 
Number of input 
neurons 
24 
Number of output 
neurons 
1 
Learning rate 0.1 
LSSVM 

 
0.292 

 
34.807 
Table 1: Parameter setting of model. 
The average absolute error is used as the fitness 
function to verify the performance of the risk prediction 
model based on LSSVM. The fitness curve of the 
enterprise credit financing risk prediction model is shown 
in Figure 5. 
 
10
11
8
9
6
7
4
5
0 50 100 150 200
Average fitness value
Optimal fitness value
MAPE (%)
Number of iterations
 
Figure 5: Fitness curve of enterprise credit financing risk 
prediction model. 
As can be seen from Figure 5, the fitness curve of 
the risk prediction model decreased rapidly in the first 52 
iterations, and the average absolute error percentage of 
the risk prediction model gradually decreased from 15.61% 
to 4.76%. The model began to converge after the 52nd 
iteration, and its average absolute error percentage was 
finally controlled at about 4.83%. In order to verify the 
effectiveness and optimization of the risk prediction 
model based on antagonistic neural network and LSSVM, 
taking the accuracy and recall of the model as the 
judgment standard, the LSSVM risk prediction model is 
compared with random forest model, support vector 
machine model and BP neural network. The accuracy 
and recall curves of the four models are shown in Figure 
6. 

150 Informatica 46 (2022) 145 –152 J. Gu 
 
SVM
BPNN
LSSVM
RF
0.8
0.6
0.4
0.2
0
1.0
0 0.2 0.4 0.6 0.8 1.0
Recall
Accuracy
 
Figure 6: Accuracy and recall curves of four models. 
As can be seen from Figure 6, the risk prediction 
model based on LSSVM has obvious advantages in 
accuracy and recall rate. The accuracy rate of the model 
is 90.15% and the recall rate is 85.63%, which is 
significantly higher than that of random forest model, 
support vector machine model and BP neural network. 
LSSVM model has obvious advantages in classification 
prediction accuracy and recall rate, and the stability of 
the model is the best, which can effectively predict and 
analyze enterprise credit financing risk. 
4.2 Prediction and empirical analysis 
In order to verify the practicability and feasibility of the 
enterprise credit financing risk prediction model based on 
LSSVM in the prediction of enterprise loan default risk, 
5000 loan enterprise user data of an enterprise loan 
platform are taken as the experimental data set to verify 
the application effect of the risk prediction model in the 
actual enterprise credit risk. Comparing the risk 
prediction model of enterprise credit financing based on 
LSSVM with random forest model, Gaussian naive 
Bayesian model and BP neural network, the root mean 
square error of the four models is shown in Figure 7. 
 
36 32 28 24 20 16 12 8 4 0
20
80
60
40
100
Number of iterations
Root mean square error
GNB BP LSSVM RF
30
90
70
50
40
 
Figure 7: Root mean square error of four models. 
As can be seen from Figure 7, the root mean square 
error of BP neural network is the largest, while the root 
mean square error of the risk prediction model based on 
LS VM is the smallest, which can effectively and 
accurately predict the enterprise credit financing risk. 
From the change trend of root mean square error curve, it 
can be seen that the risk prediction model based on 
LSSVM has the best stability, and random forest model, 
Gaussian naive Bayes model and BP neural network all 
have varying degrees of fluctuation. With the increase of 
iteration times, the root mean square error curve of the 
three models has poor stability and is prone to large-scale 
jump. The ROC change curve of the enterprise credit 
financing risk prediction model based on LSSVM is 
shown in Figure 8. 
 
Specificity
0 0.2 0.4 0.6 0.8 1.0
0
0.2
0.4
0.6
0.8
1.0
Sensitivity
SVM
BPNN
LSSVM
RF
 
Figure 8: ROC change curve of enterprise credit 
financing risk prediction model. 
As can be seen from figure 8, the area under the 
ROC curve of the enterprise credit financing risk 
prediction model based on LSSVM is 0.9677, and its 
standard deviation is 0.0023, which proves that the risk 
prediction model has a high accuracy in the classification 
and prediction of enterprise loan default. Compared with 
random forest model, Gaussian naive Bayesian model 
and BP neural network, the area under ROC curve of risk 
prediction model based on LSSVM is the largest and the 
area of BP neural network is the smallest, which proves 
the superiority of risk prediction model based on LSSVM 
in the four models. The risk prediction model based on 
LSSVM, random forest model, Gaussian naive Bayesian 
model and BP neural network are used to conduct 10 
random experiments on enterprise loan default prediction. 
The prediction error rate of the four models in the face of 
enterprise loan default and non-default is shown in 
Figure 9. 
 

Risk Prediction of Enterprise Credit Financing Management … Informatica 46 (2022) 145 –152 151 
10 9 8 7 6 5 4 3 2 1
0
18
14
10
6
4
2
16
12
8
Sample number
Default error rate （% ）
10 9 8 7 6 5 4 3 2 1
0
14
10
6
4
2
12
8
Sample number
Non default error rate （% ）
GNB BP LSSVM RF
GNB BP LSSVM RF
(b) Forecast results of non default
(a) Forecast results of default
 
Figure 9: The prediction error rate of the four models in 
the face of enterprise loan default and non-default. 
As can be seen from Figure 9, the average error rate 
of default prediction of risk prediction model based on 
LSSVM is 6.48%, while the average error rates of default 
prediction of random forest model, Gaussian naive 
Bayesian model and BP neural network are 8.25%, 12.57% 
and 15.49% respectively, which proves the superiority of 
risk prediction model based on LSSVM in enterprise 
loan default prediction, it can effectively and accurately 
predict the credit financing risk of enterprises. In the 
prediction of non-default of enterprise loans, the average 
error rate of non-default prediction of the risk prediction 
model based on LSSVM is 3.09%, which is significantly 
lower than 5.66% of the random forest model and 8.24% 
of the Gaussian naive Bayesian model. The error rate of 
non-default prediction of BP neural network is the 
highest, with an error rate of 11.81%, which proves the 
superiority of the prediction performance of the risk 
prediction model based on LSSVM. 
5 Conclusion 
Aiming at the credit financing risk control of small, 
medium-sized and micro enterprises, this paper studies 
the construction of enterprise credit financing risk 
prediction model based on antagonistic neural network 
and least squares support vector machine algorithm, uses 
antagonistic neural network to solve the problem of data 
sample imbalance, improves the richness of data samples, 
and uses least squares support vector machine model to 
predict enterprise credit financing risk. The experimental 
results of performance verification show that the 
accuracy and recall of the risk prediction model based on 
anti-neural network and branch least squares support 
vector machine are 90.15% and 85.63% respectively. In 
the actual risk prediction, the prediction error rates of the 
model in the case of default and non-default are 6.48% 
and 3.09% respectively. The prediction accuracy of 
enterprise credit risk is high, which can effectively 
predict and warn the enterprise credit default risk and 
help the enterprise carry out risk management. The 
research uses the generated neural network to generate 
false samples, which can further optimize the parameters 
of the generated anti neural network in the future, 
improve the learning performance of the model to the 
original data samples, and enhance the applicability of 
the generated samples. 
References 
[1] Stevenson M P, Mues C, Bravo C (2021). The value 
of text for small business default prediction: A 
Deep Learning approach.European Journal of 
Operational Research,295, pp. 758-771. 
https://doi.org/10.1016/j.ejor.2021.03.008 
[2] Sariannidis N, Papadakis S, Garefalakis A, 
Lemonakis C, Kyriaki-Argyro T (2020). Default 
avoidance on credit card portfolios using 
accounting, demographical and exploratory factors: 
decision making based on machine learning (ML) 
techniques.Annals of Operations Research,294, pp. 
715-739. https://doi.org/10.1007/s10479-019-
03188-0 
[3] García V, Marqués AI, Sánchez JS (2019). 
Exploring the synergetic effects of sample types on 
the performance of ensembles for credit risk and 
corporate bankruptcy prediction.Information Fusion, 
47, pp. 88-101. 
https://doi.org/10.1016/j.inffus.2018.07.004 
[4] Moscato V, Picariello A, Sperlí G (2021). A 
benchmark of machine learning approaches for 
credit score prediction.Expert Systems with 
Applications,165, pp. 113986. 
https://doi.org/10.1016/j.eswa.2020.113986 
[5] Niu K, Zhang Z, Liu Y, Li R (2020). Resampling 
Ensemble Model Based on Data Distribution for 
Imbalanced Credit Risk Evaluation in P2P 
Lending.Information Sciences,536, pp. 120-134. 
https://doi.org/10.1016/j.ins.2020.05.040 
[6] Li B, Xiao B, Yang Y (2021). Strengthen credit 
scoring system of small and micro businesses with 
soft information: Analysis and comparison based on 
neural network models.Journal of Intelligent and 
Fuzzy Systems, 40, pp. 1-18. 
https://doi.org/10.3233/JIFS-200866 
[7] Xu Y, Liu Z, Li Y, Hou H, Cao Y, Zhao Y, Guo W, 
Cui L (2020). Feature data processing: Making 

152 Informatica 46 (2022) 145 –152 J. Gu 
medical data fit deep neural networks.Future 
Generation Computer Systems,109, pp. 149-157. 
https://doi.org/10.1016/j.future.2020.02.034 
[8] Li H, Wang S , He P, de Rezende Rocha A (2020). 
Face Anti-Spoofing with Deep Neural Network 
Distillation.IEEE Journal of Selected Topics in 
Signal Processing,14, pp. 933-946. 
https://doi.org/10.1109/JSTSP.2020.3001719 
[9] Junior F, Yen G G (2021). Pruning of Generative 
Adversarial Neural Networks for Medical Imaging 
Diagnostics with Evolution Strategy.Information 
Sciences,558, pp. 91-102. 
https://doi.org/10.1016/j.ins.2020.12.086 
[10] Green A J, Mohlenkamp M J, Das J, Chaudhari M, 
Truong L, Tanguay RL, Reif DM (2021). 
Leveraging high-throughput screening data, deep 
neural networks, and conditional generative 
adversarial networks to advance predictive 
toxicology.PLOS Computational Biology,17, pp. 1-
16. https://doi.org/10.1371/journal.pcbi.1009135 
[11] Tian C, Han X (2021). A study on ship collision 
conflict prediction in the Taiwan Strait using the 
EMD-based LSSVM method.PLOS ONE,16, pp. 1-
16. https://doi.org/10.1371/journal.pone.0250948 
[12] Zhang X, Ge Z (2019). Local Parameter 
Optimization of LSSVM for Industrial Soft Sensing 
with Big Data and Cloud Implementation.IEEE 
Transactions on Industrial Informatics,16, pp. 
2917-2928. 
https://doi.org/10.1109/TII.2019.2900479 
[13] Ge Q, Guo C, Jiang H, Lu Z, Yao G, Zhang J, Hua 
Q (2020). Industrial Power Load Forecasting 
Method Based on Reinforcement Learning and 
PSO-LSSVM.IEEE Transactions on Cybernetics,52, 
pp. 1112-1124. 
https://doi.org/10.1109/TCYB.2020.2983871 
[14] Zhen C, Liu H (2019). Model for Predicting NOx 
Emission from Boilers Based on MWOA-LSSVM 
Integration.Journal of Chemical Engineering of 
Japan,52, pp. 702-709. 
https://doi.org/10.1252/jcej.19we004 
[15] Zhao Z, Chen K, Chen Y, Dai Y, Liu Z, Zhao K, 
Wang H, Peng Z (2021). An Ultra-Fast Power 
Prediction Method Based on Simplified LSSVM 
Hyperparameters Optimization for PV Power 
Smoothing.Energies,14, pp. 1-14. 
https://doi.org/10.3390/en14185752