Research ArticleResearch on Application of Big Data in Internet Financial CreditInvestigation Based on Improved GA-BP Neural Network

Fei-Peng Wang

School of Business Administration, Shandong Institute of Business and Technology, Yantai, Shandong 264005, China

Correspondence should be addressed to Fei-Peng Wang; wfp2009@sdibt.edu.cn

Received 20 June 2018; Revised 11 September 2018; Accepted 18 September 2018; Published 2 December 2018

Guest Editor: Zhihan Lv

Copyright © 2018 Fei-Peng Wang. This is an open access article distributed under the Creative Commons Attribution License,which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

The arrival of the era of big data has provided a new direction of development for internet financial credit collection. First of all, thearticle introduced the situation of internet finance and traditional credit industry. Based on that, the mathematical model was usedto demonstrate the necessity of developing big data financial credit information. Then, the Internet financial credit data arepreprocessed, the variables suitable for modeling are selected, and the dynamic credit tracking model of BP neural networkbased on adaptive genetic algorithm is constructed. It is found that both LM training algorithm and Bayesian algorithm canconverge the error to 10e-6 quickly in the model training, and the overall training effect is ideal. Finally, the rule extractionalgorithm is used to simulate the test samples. The accuracy rate of each sample method is over 90%, and some accuracy rate iseven more than 90%, which indicates that the model is applicable to the credit data of big data in internet finance.

1. Introduction

The credit system is the cornerstone to the development ofthe market economy and the financial industry. A soundcredit system helps create a good consumer environment,effectively prevents the spread of credit risks, and promotesthe healthy development of the economy [1]. Internet financehas been regarded by more and more scholars as the newengine of economic growth, and relying on big data to estab-lish the credit system has become an inevitable choice for thedevelopment of internet finance.

Many researchers have put forward some creative ideasfor the credit reporting system of big data [2]. However,due to the fact that large data credit is a brand-new concept,there are relatively few literatures directly related to it, andthe related research lacks systematisms and depth. Only thewhole idea of building a big data credit system was explored[3], and there was no reasonable evidence for feasibility anal-ysis. In view of this, under the background of big data ofInternet finance and domestic traditional credit investigationindustry, the construction of personal credit investigationsystem is discussed. The development of the research ismainly reflected in three aspects. First, based on economic

theory, concepts of credit information costs and potentialrisk costs are raised and the use of quantitative models todemonstrate the necessity of developing big data credits.Second, ingeniously introduce conducts and contacts ofInternet data into BP neural network algorithms based onadaptive genetic algorithms. A credit-tracking model basedon big data was constructed. Third, the application of themodel was briefly described.

2. Internet Finance and Credit Industry

2.1. Internet Finance. Finance is at the core of modern eco-nomic operations, and financial viability determines thequality and potential of the overall economy. Financialdynamism depends on the ability of the financial system toaccept new concepts and apply new technologies. However,with the application and popularization of Internet technol-ogy in the financial field, the traditional finance has beenrapidly transformed from the economic field into a newindustry—Internet finance [4]. Internet finance refers to thetraditional financial institutions and Internet enterprisesusing Internet technology and information and communica-tion technology to achieve financing, payment, investment,

HindawiComplexityVolume 2018, Article ID 7616537, 16 pageshttps://doi.org/10.1155/2018/7616537

and information intermediary services. It is not only differentfrom the indirect financing of commercial banks but alsodifferent from the direct financing of the capital marketinformation finance business model. Internet finance is anew type of business generated by the cross-border integra-tion of the traditional financial industry with the Internetindustry with new technologies such as cloud computing,big data, and mobile payment [5].

2.2. Traditional Credit Investigation Industry. The essenceof internet finance is finance, and the core of finance iscredit. Therefore, the credit information system is an effec-tive constraint mechanism to ensure the integrity of thefinancial market.

The traditional personal credit system in the financialindustry is inefficient, narrow coverage, and slow update ofdata, which cannot satisfy the control of personal credit riskof Internet finance. At present, China’s credit reportingsystem is mainly represented by the People’s Bank ofChina Credit Reporting Center; there are also local govern-ment and its functional departments led by the governmentcredit reporting system, as well as the market-orientedcredit reporting system represented by the eight creditreporting companies that have obtained the licenses ofindividual credit institutions [6], as shown in Figure 1.At the same time, China also issued some more authorita-tive regulations and regulations on the establishment of acredit information system.

3. Mathematical Models

Based on the knowledge of economics, the concepts of creditcosts and potential risk costs are proposed in this paper.Starting from the three characteristics of big data credit,using mathematical models to demonstrate the necessity ofdeveloping big data personal credit is put forward.

3.1. Analysis of Idea. Compared with the superiority of tradi-tional credit information, big data credit can be measuredfrom two dimensions of efficiency and cost [7]. Because ofthe three characteristics of timeliness, accuracy, and econo-mies of scale in big data credits, the study believes that bigdata credits are superior to traditional credit information interms of efficiency [8]. This paper will focus on the field of

personal credit reporting from the cost dimension and thendemonstrate the necessity of developing large data credit.The reasoning argument is based on the three characteristicsof big data credits. By setting assumptions and constructing amathematical model, the total social costs incurred by thetwo are compared.

The cost of traditional credit information mainlyincludes credit cost, potential risk cost, and other costs [9].Firstly, each additional coverage of the credit system willresult in the corresponding operating costs, which arederived from the data acquisition, data analysis, personnelemployment, and other necessary aspects of credit, anddefined as the cost of credit; secondly, even if the creditbureaus do not continue to expand their credit coverage, theystill need to provide services to the creditworthy groups andpay for necessary equipment maintenance costs. This partof the cost is defined as other costs. Finally, based on eco-nomic rational people assumptions, any information subjectis a potential defaulter. The purpose of credit investigation isto prevent the credit risk brought about by breach of con-tract. Conversely, without credit checking, potential riskscan be generated, which is called potential risk cost. Creditevaluation of information subjects through credit informa-tion can effectively reduce potential risk costs but cannoteliminate them.

As a brand-new information network technology, largedata credit reporting technology research and develop-ment, supporting infrastructure is needed. Therefore, largedata credit has three costs of traditional credit, but at thesame time, it also needs technical input, that is, technicalcost [10].

3.2. Assumed Conditions and Variable Settings. For the sakeof discussion, set the following six assumptions [11].

First, the information subject is homogenous, and there isno difference in the potential risk cost caused to the creditsystem. Assuming that the potential risk cost per unit timeof the nonaccepting information subject is p, and the poten-tial risk cost per unit time after accepting the traditionalcredit and large data credit is p1 andp2, respectively.

Second, the current population N is a fairly large value.Both the big data credit and the traditional credit informa-tion cannot include the entire population within the scopeof the credit investigation within the time frame studied bythe institute and do not consider the population growth.

Thirdly, the speed of credit information is defined as thenumber of people newly covered by the credit system in agiven period of time. It is assumed that once an informationsubject has been included in the credit system, it will not needto reconduct it, and the traditional credit and big data creditsare m1 andm1, respectively.

Fourth, the current credit system coverage is 0, and it isassumed that big data credits and traditional credits havethe same credit cost β when the coverage is zero.

Fifth, big data credits need to invest a fixed technical costA per unit of time, regardless of the diminishing cost of tech-nology brought about by the maturity of big data credits.

Sixth, big data credit and traditional credit have the sameother cost R T .

10%

Pursuit new credit

10%Credit mix

Outstanding debt30%35%

15%

Lit history length

Payment history

Figure 1: Reference elements of the traditional credit scoringmodel.

2 Complexity

Based on the three characteristics of big data credit, thispaper analyzes the difference in cost between big data creditand traditional credit information [12].

First, due to the timeliness of big data credits, the studyconsiders that big data credits have higher credit reportingspeeds than traditional credit information, that is, m1 <m2.

Second, because the assessment results of big data creditsare more accurate, the study considers that big data creditscan reduce the potential risk cost more effectively than tradi-tional credit information, that is, p1 > p2.

Thirdly, due to the scale effect of big data credits, thestudy considers that the credit cost of big data credits doesnot increase with the expansion of credit coverage, that is,the credit cost β of the unit population remains unchanged.The credit information cost of the traditional credit reportingunit population will increase with the increase in the numberof people covered. For ease of discussion, the article sets thegrowth of its credit cost as a linear model n − 1 α + β [13],where n is the number of people covered by the currentcredit system.

3.3. Model Construction and Solution. The study will use timeT as an independent variable and set the current time pointas T = 0 to analyze the total social cost accumulated on0, T for traditional credit andbigdata credits.When the timeisT , the populations covered by traditional credit and big datacredits are m1T and m2T , respectively. And the uncoveredpopulations areN −m1T andN −m2T , respectively.

The total social cost of traditional credit information on0, T is C1 T , which is the sum of credit cost, potential riskcost, and other costs:

C1 T = 〠m1t

n=1n − 1 α + β +

T

0p N −m1t dt +

T

0p1m1tdt + R T

1

The total social cost C2 T of 0, T for big data credit isthe sum of credit cost, potential risk cost, other cost, andtechnical cost:

C2 =m2Tβ +T

0p N −m2t dt +

T

0p2m2tdt + R T + AT

2

For all T , there are

C T = C1 T − C2 T

= 〠m1t

n=1n − 1 α + m1T −m2T β +

T

0p m2t −m1t dt

+T

0p1m1 − p2m2 tdt − AT

3

Finding the first and second derivatives of C T yields

C′ T =m21Tα + m2 p − p2 −m1 p − p1 T

+ m1 −m2 β −m1α

2 − A,4

C″ T =m21α + m2 p − p2 −m1 p − p1 5

Since m1 <m2 and p1 > p2 > p3, C″ T > 0 is obtained,and the derivative C″ T monotonously increases in theinterval 0, T , and C″ T < 0. Judging that C T shows atrend of decreasing first and then increasing, there is a mini-mum value and no maximum value. Based on the assump-tion of N infinity, there must be a bit of TE in the intervalof 0, T , makingC T = 0.

Let C1 T = C2 T , get

TE =2A + m1 −m2 β

m21 + m2 p − p2 −m1 p − p1

> 0 6

When T < TE , C T < 0, that is, the total social cost ofcrediting using big data is greater than the total social costof traditional credit information; and when T > TE , the totalsocial cost of credit using big data is less than the total socialcost of traditional credit information.

Defining T < TE and T > TE as short term and longterm, the following conclusions are drawn: in the shortterm, due to the high technical cost of investing in big datacredits, traditional credit has more advantages in terms ofcost; but in the long run, large data credit has three charac-teristics: timeliness, accuracy, and economies of scale.Whether from the perspective of efficiency or cost, largedata credit is superior to traditional credit. Therefore, inthe theoretical level, it is necessary to develop large datacredit reporting.

4. Management of Big Data Based on InternetFinancial Credits

After demonstrating the necessity of big data credit, thecredit tracking model based on adaptive GA-BP neural net-work will be studied. Before this, first explain big data anddo preprocessing.

4.1. Credit Market Big Data Overview. The unbalanced open-ing of the central bank’s credit system still remains unsolved.At least the central bank’s attitude is clear and supports thedevelopment of internet finance, and it is believed that inter-net finance is a useful complement to traditional finance [14].Most of them leave credit data in the databases of other insti-tutions outside the bank credit system and Internet compa-nies. Internet credit companies have a strong demand forcredit ratings for lenders’ credit ratings. The market sponta-neously formed a unique risk-control ecological field. Largecompanies use self-built credit rating systems through datamining; small companies obtain credit rating consultingservices through third-party information sharing.

3Complexity

4.2. The Source of Big Data. E-commerce big data are used forrisk control. After all the information is aggregated, thevalues are entered into a network behavior scoring modelfor credit rating. Big data on credit card websites is also veryvaluable for risk control of Internet finance. The year ofcredit card application, whether it passes, credit line, cardtype, credit card repayment amount, attention to preferentialinformation, etc. can be used as reference data for credit rat-ing: use social network relationship data and mutual trustbetween friends to aggregate popularity [15].

The borrower is divided into several credit ratings, butit is not necessary to publish their own credit history. Inaddition, water and electricity bill payment informationin Taobao, credit card repayment information, payment,and transaction information have become data all-roundplayers: the credit big data includes credit limits anddefault records. The direction of third-party payment plat-form payment, the amount of monthly payments, and thepurchase of product brands can all be used as importantreference data for credit rating: big data such as waterfor life service websites. Electricity, gas, cable television,telephone, network fees, and property fee payment plat-forms objectively and truly reflect the basic informationof individuals and are an important type of data in creditrating, as shown in Figure 2.

4.3. Internet Credit Data Preprocessing. Through the con-struction of the user’s credit image, it is possible to rationallyorganize and store the dimensions of the complex, diverse,widely distributed, and heterogeneous platforms on theInternet. However, big data are sparsely populated andonline and offline behaviors of users are widely distributedand extremely difficult. Full collection and coverage and userbehavior preferences are also different; there are significantdifferences in the behavior of different categories, resulting

in the possibility of user behavior information missing rateof more than 50%, plus the data source instability causedby the lack of data and inconsistencies. The problem hascaused us to preprocess large data before using it to modelfinancial information, ensuring that the data meet the model-ing requirements. Pretreatment mainly starts from thefollowing aspects.

4.3.1. Data Cleaning. The purpose of data cleaning is mainlyto deal with the data problems found in the process of dataverification. Its purpose is to solve the problems of incom-plete data, inconsistent data, and data noise. During the datacleaning process, the data are properly processed andadjusted for problematic data so that it can meet the require-ments of modeling as much as possible because the quality ofthe model largely depends on the amount of modeling data; ifthe adjusted data are still unavailable, they need to be deleted[16]. In the specific cleaning process, the uniqueness, com-pleteness, validity, relevance, timeliness, and consistencymust be ensured.

The binning method is a commonly used method inthe data cleaning process. Its core idea is to smooth thevalue of the current data by the value of the surroundingdata. There are mainly three methods for data smoothing,as shown in Table 1.

Threedata smoothingmethods and specific steps are givenin Table 1, so we can choose the appropriate data smoothingmethod according to the characteristics of the data.

4.3.2. Univariate Analysis. The purpose of univariate analysisis to determine the variables that satisfy the following twoconditions [17]:

(1) Conforming to actual business significance

(2) High discrimination ability for the analysis object

Risk control related big data

E-commerce website big data

Social website big data

Small loan website big data

Payment website big data

Life service website big data

Representing a company or product

Ali, Jingdong, Suning

Silver rate network

Sina Weibo, Tencent Wechat

Peer-to-peer lending

Alipay Yi Bao

Ping An Account

Credit card website big data

Figure 2: Data sources.

Table 1: Summary of data smoothing methods.

Method name Specific steps

Smooth by average Average the data in the box and replace all values in the box with the average

Smoothed by boundary values Replacing each value in the box with the smallest boundary value

By median smoothing Take the median of the box to replace all the data in the box

4 Complexity

The results obtained through univariate analysis are a setof variables that are basically suitable for modeling, whilereducing the complexity of later multivariate analysis.

At this stage, usually from the analysis of the ability todistinguish the variables, the following method is used toanalyze the variables several times, until the variables meetthe above requirements, as shown in Figure 3:

4.3.3. Multivariate Analysis. Through multivariate analysis,the combination of model variables that meet the followingthree conditions can be determined:

(1) Low correlation between variables

(2) The model has stable and high discrimination ability

(3) The model contains as many different types of infor-mation as possible

4.3.4. Big Data Processing Flow. The first stage of the processis shown in Figure 4. Through the steps of data collection,data cleaning, and variable analysis, the user’s portrait dataare converted to data available for risk assessment modelingof large data; then, the effectiveness of univariate analysisand cross-variables is performed through binning, and then,multivariate analysis is performed. According to the princi-ples of comprehensiveness, scientific, feasibility, and measur-ability, the variables suitable for modeling are selected andthe BP neural network-based credit tracking model is intro-duced, as shown in Figure 4.

So, a tracking indicator system was constructed, as shownin Table 2.

According to expert experience, knowledge, and intuition,initial values of various indicators are obtained in Table 2.

5. A Dynamic Credit Tracking Model of BPNeural Network Based on AdaptiveGenetic Algorithm

Firstly, a dynamic credit tracking model was built. Then, themodel was simulated using the MATLAB neural networktoolbox. Finally, the model was used to train 20 samples col-lected through the network using LM training algorithm,Bias algorithm, and momentum gradient algorithm, and theremaining 8 samples were simulated. There are three reasonsfor the number of samples selected:

(1) There are many indicators, and the horizontal dataare more

(2) The data preprocessing is reduced by a part

(3) The amount of data is large

However, only 20 samples were selected for analysis dueto objective reasons. Very good simulation results wereachieved, as shown in the flowchart in Figure 5.

As a powerful tool for studying complexity, BP (BackPropagation) neural network technology has demonstratedextraordinary advantages in pattern recognition, classifica-tion, prediction, and rating in recent years. It has a powerfulparallel processing mechanism and is highly self-learningand self-developed. Adaptability, there are a large numberof adjustable parameters inside, thus making the systemmore flexible and able to handle any type of data, which isunmatched by many traditional methods. Through continu-ous learning, neural networks can discover its laws from alarge amount of complex data of unknown patterns. It over-comes the complexity of the traditional analysis process andthe difficulty of selecting the appropriate model functionform. It is a natural nonlinear modeling process. It is neces-sary to distinguish what kind of nonlinear relationshipexisted and brought great results to modeling and analysis.When the method is applied to credit risk analysis, the inputof a series of popular credit indexes can be processed, and thecorresponding credit rating output can be produced, and theexperience, knowledge, and intuitive thinking of experts canbe reproduced, thus ensuring the objectivity of the evaluationand prediction results. Because the initial weights of the BPneural network are randomly generated and the trainingspeed is slow, there are problems such as local minimumvalues [18]. Therefore, these defects can be improved to someextent through the combination of genetic optimization andBP neural networks.

5.1. Dynamic Tracking Model Index System Construction.Basedon theprinciples of comprehensiveness, scientific, feasi-bility, and testability, after preprocessing the data, this paperfinally selected two first-level indicators and a total of 11second-level indicators to construct a dynamic credit trackingevaluation index system. The design steps are as follows.

Firstly, the initial indicators were determined. Throughdrawing lessons from FICO’s credit scoring index system inthe United States, the personal credit evaluation index systemof commercial banks, and the actual experience and expertopinions of China’s public credit management, this paperobtained a number of preliminary indicators that reflect thepublic credit rating and then classify the goals by connotation.It is a set of three key indicator sets, that is, a first-levelindicator. Each indicator set contains multiple secondaryindicators that reflect its connotation and are operational.

Secondly, according to the three principles of the sameindicators with the same connotation to be merged, the causeand effect of the indicators of cause and effect, and poor oper-ability to find alternative indicators, the indicators werescreened by the Delphi expert assessment method and relatedanalysis methods; for example, we removed the obstacles

Economicimplication

analyis

Variableconversion

Variablediscriminationability analysis

Figure 3: Univariate analysis.

5Complexity

such as “asset-liabilities ratio” and “ineffectiveness” and thenremoved the indicators “vehicles” and “blacklist of personalcredit information.” The presurvey shows that the indicatorhas strong homogeneity. Finally, it simplifies the indicator“personal annual expenditure,” and the correlation analysisshows that it is strongly related to “U21 Contact the bank.”

Thirdly, expert circular appraisal and revision are per-formed on the selected credit index to form the final creditindex system. In order to ensure the quality of the index sys-tem, we must further find relevant experts to demonstrate.Experts who are invited here should not repeat the expertswho collect and screen indicators. After the experts have dem-onstrated and revised the credit index system, they have con-ducted presurveys and solicited opinions from the surveyed

people. Such a cycle can then be determined as the final creditindex system.

Finally, determine the scoring criteria. Through theDelphi expert evaluation method, the opinions and opinionsof internal and external experts were investigated and repeat-edly integrated. Finally, they obtained the unanimouslyagreed opinions and opinions as the basis of the index scor-ing standards. For example, with the “Contact with Banks”as an example, the more contacts there are, the stronger therepayment willingness and the better the credit status, sothe “contact with the bank” is assigned 5 points in descendingorder, 3 points, 0 point is similar to other indicators areassigned in order. So, a tracking indicator system wasconstructed, as shown in Table 3.

Start

Financial credittracking model

Data collection

Data clearing

Univariate analysis

Multivariate analysisGenerate cross variables. Split

binning of crossover variables andanalyze the effectiveness of

crossover variables

Splitting categorical variables andcontinuous variables to analyzethe effectiveness of univariate

variables

Figure 4: Big data processing flow.

Table 2: Sample.

Serial number index U11 U12 …… U15 U16 U21 U22 …… U25 U26 U31 …… U34 Credit rating score

1 5 5 …… 5 5 1 1 …… 3 1 4 …… 3 2

2 3 3 …… 3 3 5 3 …… 3 3 4 …… 3 2

3 5 1 …… 3 1 5 1 …… 1 3 5 …… 3 2

…… …… …… …… …… …… …… …… …… …… …… …… …… …… ……

8 5 4 …… 5 1 3 3 …… 4 3 3 …… 5 2

9 3 4 …… 5 5 3 1 …… 5 3 4 …… 3 2

10 3 3 …… 1 5 5 1 …… 1 1 5 …… 5 1

…… …… …… …… …… …… …… …… …… …… …… …… …… …… ……

15 5 5 …… 3 5 3 1 …… 1 4 4 …… 5 2

16 5 3 …… 1 1 1 3 …… 3 3 5 …… 1 1

17 3 1 …… 3 3 5 1 …… 4 1 4 …… 5 2

…… …… …… …… …… …… …… …… …… …… …… …… …… …… ……

22 3 5 …… 5 3 3 1 …… 5 5 4 …… 5 2

23 5 1 …… 3 5 1 1 …… 4 4 3 …… 3 2

24 3 4 …… 5 5 1 1 …… 3 1 4 …… 3 2

…… …… …… …… …… …… …… …… …… …… …… …… …… …… ……

6 Complexity

5.2. Construction of Dynamic Credit Tracking Model Basedon BP Neural Network Based on Adaptive Genetic Algorithm

5.2.1. Genetic Algorithm. The genetic algorithm is an algo-rithm that simulates the rules of the survival of the fittest inthe natural world. It selects, crosses, and mutates the popula-tion to obtain the optimal individual population method. Theoptimization process of the traditional genetic algorithm is asfollows [19]:

(1) According to the characteristics of the problem to bedealt with, select the code corresponding to the prob-lem solution, and give an initial population, whichincludes N chromosomes

(2) Calculate the fitness function value for each chromo-some in the initial population

(3) When the result of an iteration of the genetic algo-rithm meets the condition of stopping the iteration,the algorithm stops iterating. Otherwise, a randomprobability value is used to randomly select Nchromosomes from the old population, and the newpopulation composed of these chromosomes isnext iteration

(4) Crossing to obtain a cross set of N chromosomes, thenew generation of individuals will inherit the previ-ous generation of information

(5) Set a small mutation probability to allow certaingenes in the chromosome to mutate, obtain new pop-ulations to enhance individual fitness, and repeat (2)the calculation process

Constructing credit tracking model(BP neural network)

Conclusions and policyrecommendations

Model simulation(based on MATLAB)

Model training(20 internet survey samples)

Model simulation(8 online survey samples)

Figure 5: Construction and application flowchart of dynamic credittracking rating model.

The setting of the crossover probability Pcand the muta-tion probability Pm in the genetic algorithm will largely affectthe convergence of the genetic algorithm and increase theerror of the optimal solution and the real solution.

In general, the greater the value of the crossover probabil-ity Pc, the faster the new individual will be produced, and thevalue of the crossover probability Pc is too large, so that theindividual structure with a high fitness value is destroyed.For the mutation probability Pm, if the value of Pm is toosmall, it is not easy to generate a new individual; if the valueof Pm is too large, the algorithm is similar to the randomsearch algorithm.

Therefore, the crossover probability Pc and the mutationprobability Pm have a profound effect on the performance ofthe algorithm. Therefore, the crossover probability andmutation probability that can be adaptively adjusted are usedto ensure the diversity of the group:

Pc =1

e −k1Δf + 1, 7

Pc = 1 − 1e −k2Δf + 1

8

Among them, Δf means the difference between the max-imum fitness value and the average fitness value of each chro-mosome individual; k1 and k2 represent the adjustment rate,and they all take a value of 1. Through the adaptive geneticalgorithm, the ability of the genetic algorithm to search theglobal optimal solution can be effectively improved, and theproblem of falling into local optimization can be avoided.

5.2.2. Neural Network. Neural networks have been widelyused in the field of bank risk management. A certain struc-ture of BP neural network has the ability to predict and clas-sify similar data after training with given sample data.Applying this feature of BP neural network, according tothe established personal credit risk tracking index system, athree-layer BP neural network model was established. Thenumber of input nodes is equal to the number of feature var-iables in the index system; the number of output nodes is one,and the level of personal credit can be determined accordingto the output value; the method of determining the hiddenlayer nodes is to determine the excessive hidden layer nodesfirst, after training, and then according to the training resultsare pruned [20].

The connection weight between nodes selects the randomnumber in the interval [−1, 1], and the optimal initial weightof the BP neural network is determined by the adaptivegenetic algorithm. The transfer function of each layer is tan-sig, and the transfer function of the second layer is purelin, asshown in Figure 6.

5.3. Dynamic Credit Tracking Model Training

5.3.1. Model Training Steps. According to the previouslyestablished tracking index system, a certain amount of per-sonal data is selected in the personal credit database, eachof which contains a corresponding personal credit level L,represented by the vector Aj = a1, a2,… , an, d , where ai is

7Complexity

Table3:Trackingindicatorsystem

.

First-levelind

icators

Second

aryindicators

Indexscore

U1person

albasicinform

ation

U11

sex

Male

5Female

3

U12

age

Ages23

to27

3Ages28

to32

432

yearsof

ageor

older

5Age

22andbelow

1

U13

educationlevel

Junior

college

andbelow

1Und

ergraduatecourse

3Masteror

above

5

U14

marriage

Married

5Unm

arried

3Divorced

1

U15

registered

perm

anentresidence

Rural

1In

thiscity

5Other

city

3

U16

administrativelevel/po

sition

Division/generalm

anager

orabove

5Section/partmanager

3Other

1

U17

person

alannu

alincome

Over100,000yuan

530,000

to100,000yuan

410,000

yuan

to30,000

yuan

3Lessthan

10,000

yuan

1

U2person

alcreditindex

U21

contactthebank

Keepin

touch

5Frequent

contact

4Occasionally

contact

3Never

contact

1

U22

loan

bank

accoun

tstatus

Havealoan

accoun

t3

Savingsaccoun

t5

Haveacreditcard

accoun

t4

There

isno

accoun

t1

U23

creditcard,m

ortgagedelin

quency

record

05

1–2times

33times

andabove

1

U24

badrecordson

taxes,utilities,etc.

05

1–2times

33times

andabove

1

8 Complexity

the corresponding feature variable in F′,ai=5, 3 or 0, n is thenumber of elements in F′, n = 11, L ∈D, D = A, B, C,D ,and n L is the position number of L in set D, 1 ≤ n L ≤ 4,d = d1, d2,… , d4 , and dn l = 1, di = 0, 1 ≤ i ≤ 4, i ≠ n Lis personal credit rating vector. In the data set, part ofthe data is randomly selected as a test data set for verifica-tion of model training results; the remaining data are usedas a training data set for model learning.

a1, a2,… , an of each vector Aj of the training data set istaken as the input of the model, and d is the target outputof the model. Let wlm be the connection weight of input layernode l to hidden layer node m, vmp be the connection weightvalue of hidden layer node l to output layer node p, and θand w are the determined threshold values. For a giveninput a1, a2,… , an, the activation function of hidden nodem is tansig, and the activation function of output node pis purelin.

When the genetic algorithm optimizes the initial weightsof the BP neural network, a fitness function needs to be set todetermine the probability that the individual is selected.Since the fitness value of the genetic algorithm is continuouslyincreased during the search process of the optimal value, thefitness function can be set as the following expression:

f itness = 1∑N

i=1 yi − xi + ζ, 9

where xi represents the actual credit score data; yi representsthe credit score data predicted by the BP neural network; Nrepresents the number of samples; ζ is a smaller value, avoid-ing 0 as the denominator.

Then, according to the genetic algorithm optimizationstep introduced, the initial weight of the BP neural networkis determined.

For each sample data in the training data set, if

maxp

βp − dp ≤ η1 maxp

βp − dp ≤ η1, 10

then the training data set can be correctly classified,where 0 < η1 < 1/2. The test data set is used to testthe trained model. If (10) is satisfied, the model train-ing is completed; otherwise, the model is trained again.The purpose of model training is to obtain a set ofweights w, v that can correctly classify personal creditrating data.

After the model training is completed, a fully connectedthree-layer BP network is obtained. The model pruning isto delete some of the connections in the model according tocertain rules without affecting the accuracy of the model clas-sification. In the weighted set w, v of the trained model,under the condition that (10) is satisfied, for any wlm, ifit satisfies

maxp

vmp − vlp ≤ 4η2, 11

then delete wlm in the model. For any vmp, if satisfied

maxp

vmp ≤ 4η2, 12

then delete vmp in the model, where η2 > 0 and η1 + η2 <1/2. If there is no weight that satisfies (11) and (12) inthe weight set w, v , ωlm associated with the vectorminimum product max

pvmp − vlp is deleted in the weight

set w, v .After the model training is completed, MATLAB will

automatically extract the model rules. The rule extraction isto extract the classification rules in the pruned model, and

x1

x2

xn

.

.

.

.

.

.

Y

Calculation process

Learningprocess

Error calculationevaluation

Enter Expectedoutput

Output

Output layerHidden layerInput layer

Weight correction

w1

w2

wn

Figure 6: Structure of credit tracking model based on BP neural network.

9Complexity

the relationship between the input and output of the modelperformance after the pruning is still more complicated.The rule extraction algorithm firstly uses a clusteringmethod to discretely process the hidden layer activationvalue. When certain accuracy of the model is ensured, theinput value and the hidden layer activation value are discre-tized, and the number of discrete values can be conve-niently managed. Second, enumerate the activation valuesof the discretization, calculate the output of the model,and generate complete rules from the hidden layer activationvalue to the output. Third, for each hidden layer activationvalue that appears in the above rules, the enumeration cangenerate these hidden layer activations. Enter the value andgenerate a complete rule from input to output, as shown inFigure 7.

Using the above adaptive GA-BP neural network model,it is possible to dynamically track the implementation ofcredit rating changes and obtain the following credit ratingtransition table, as shown in Table 4.

The credit rating is divided into four levels: A, B, C, andD. When the probability of default does not exceed 5%, wedetermine the risk classification of “Normal.” When theprobability of repayment does not exceed 5%, we determinethe risk classification of “Track.”

5.3.2. Result Analysis. After completing the design of themodel, the model was simulated by the MATLAB program.20 samples were used for training. The results are shown inFigures 8, 9, and 10. Both the LM training algorithm andthe Bayesian algorithm can converge to 10e-6, but momen-tum gradient algorithm convergence is slow, but overall, themodel has achieved a good training effect and has a certainapplication significance.

(1) LM Training Results. It can be seen from Figure 8 that thetarget error 10e-6 is reached after five network training times;the training speed is fast, and the training effect is good.

(2) Bayesian Training Results. It can be seen from Figure 8that the target error 10e-6 is reached after 77 network train-ing times, and the training speed is fast and the training effectis good.

(3) Momentum Gradient Training Results. From Figure 10, itcan be seen that after the number of network trainings of4381 times, the target error is 10e-2 and the training speedis slow, but the training result is still within the allowableerror range.

5.4. Dynamic Credit Tracking Model Simulation. To test thedynamic credit tracking model established above, theremaining 8 test samples were input into the trained networkfor testing. The network test output results are shown inFigures 11, 12, and 13. Comparing the output of the networkwith the expected output, it can be seen that the error is lessthan the allowable error. Therefore, it is considered that thenetwork output is reasonable, and the trained network hasbetter generalization performance.

After simulation, the simulation results obtained areshown in Table 5. The simulation accuracy rate of each sam-ple method has reached more than 80%, and some evenreached more than 90%. The simulation results are verygood. It can be concluded that the evaluation model is feasi-ble for comprehensive credit evaluation. This model can bedirectly used to rating unknown samples, thus reducing theevaluation workload, reducing the subjectivity of evaluation,and improving the rationality of the rating results. The

Start

Initialize BP neural network weight andgenetic algorithm parameters

Get BP neural network optimal weight

Output results

End

Genetic algorithm individual decoding

BP neural network training obtainserror as fitness value

Adaptive genetic algorithm operation

Calculate fitness value

Satisfy theend condition

Calculation error

Weight update

Satisfy theend condition

Figure 7: Flow diagram of BP neural network algorithm based on optimization of adaptive genetic algorithm.

10 Complexity

dynamic concept in the dynamic credit tracking model isreflected in the time dimension. The index data of each per-son are different in different periods, so the results predictedby the model evaluation are different. The results are dis-played in time series to reflect the dynamic concept.

As can be seen from Table 5, the simulation accuracy ofeach sample method is over 80%, and some even reach100%. The simulation results are very good.

5.5. Application of GA-BP Neural Network Model. Themicro-level dynamic credit tracking model corrects the informationasymmetry of the loan, thereby regulating the behavior of therelevant market entities. At the macro level, it is the servicesupervision that promotes financial stability. Specifically,the credit tracking model mainly achieves the following fivemajor functions: (1) automatically identify the credit risk ofthe loan; (2) dynamically monitor the credit status of thelender; (3) automatically alert the credit risk; (4) provide riskanalysis report and risk operation suggestion; and (5) reportsand risk operational advice.

The credit status of the lender has changed, and the ratingsystem needs to be rerated to reflect the true credit level in atimely manner. A personal credit rating system based on thelender’s personal basic information indicators including per-sonal economic ability indicators and personal credit indica-tors is established. The creditor’s credit file is establishedbased on the results of the credit rating in the personal creditrating system. These files also contain subfiles in the modulesthat will be mentioned below. Since the credit rating weightand indicator scores in the personal credit rating system needto be dynamically adjusted according to the changes of thelender’s actual situation, the credit status in the personalcredit file is constantly changing.

In this paper, strict postloan risk monitoring was imple-mented through the credit tracking model, and its functionalstructure was divided into four modules according todifferent degrees of loan expected default probability and

Table 4: Credit rating shift table.

Change in credit rating Risk classification Risk definition

A Normal Repayments may be stable and will not be breached. Probability of default will not exceed 5%.

B Attention Repayment may have certain doubts, default probability is generally between 5% and 20%.

C Focus on Repayment may have big doubts, the probability of default is roughly between 20%~95%.D Track Repayment may be minimal, no more than 5%.

0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 50

1

Performance is 2.48091e-009.Goal is 1e-006

5 epochs

Trai

ning

-blu

e Goa

l-bla

ck

10−8

10−6

10−4

10−2

Figure 8: LM training results.

0

1

Training SSE = 6.18657e-007

Tr-b

lue

0

1Squared weights = 3.63663

SSW

0 10 20 30 40 50 60 700

100200300400

Effective number of parameters = 17.9902

77 epochs

# pa

ram

eter

s

Figure 9: Bayesian training results.

11Complexity

corresponding credit rating of the lender, namely contactmodule, tracking and supervision module, warning module,and disciplinary module. These four submodules are at dif-ferent expected default probability and risk level, whichmay exist at the same time, and may exist in turn as the risklevel changes with the extension of the loan period. Thespecific design is as follows.

5.5.1. Contact the Module. The credit rating is attributed tothe contact module in the normal risk category, indicatingthat the expected default probability is low, the credit ratingis increased, and the loan repayment may be stable. In thecontact module, a contact interval based on the probabilityof default is set and issue different warning signals accordingto different levels of expected default probability. Based on

this, the bank establishes a contact subfile in the credit fileand maintains different degrees of contact with the lenderaccordingly, supervising its timely repayment during thecredit period. At the same time, the discount on the repay-ment can be appropriately given.

5.5.2. Tracking and Urging Modules. Concerned about therisk level in the personal credit rating system, a trackingand supervision interval based on the correspondingexpected default probability is set. In the credit trackingmodel, the expected default tracking index is determined,that is, the best default probability under the bank utilitymaximization (expected default loss and income equilibriumstate), and the default probability is used as a standard todivide the three subintervals of high, medium, and low to

0 500 1000 1500 2000 2500 3000 3500 4000

Perfomance is 0.00998595,Goal is 0.01

4381 epochs

Trai

ning

-blu

e Goa

l-bla

ck

10−3

10−1

100

101

10−2

Figure 10: Momentum gradient training results.

0 1 2 3 4 5 6 7 8Sample number

0

0.5

1

1.5

2

2.5

3

3.5

4

4.5

Cred

it ra

ting

scor

e

L-M algorithm simulation

Simulation dataRaw data

Figure 11: LM test results.

0 1 2 3 4 5 6 7 8Sample number

0

0.5

1

1.5

2

2.5

3

3.5

4

4.5Cr

edit

ratin

g sc

ore

Bayesian algorithm simulation results

Simulation dataRaw data

Figure 12: Bayes test results.

12 Complexity

reflect the change in credit status. Banks are advised to adoptdifferent tracking strategies. At the same time, a trackingsubfile is established in the credit file to record the credit sta-tus of such lenders. For lenders who do not contact the bankon time or do not contact the bank, they should reevaluatetheir credit level, pay close attention to and promptly sendthem a warning signal, or even urge them to repay inadvance. Lenders who do have difficulties in repaying loansshould be appropriately extended or reduced in order topromote their repayment enthusiasm and help to improvetheir credit level.

5.5.3. Early Warning Module. With the extension of therepayment period, the credit status of the lenders is con-stantly changing. The data analysis and evaluation in theabove files are combined to classify the lenders whose risklevel is on the suspicious level to establish an early warningsubfile to record their credit during the warning period. Sim-ilarly, an early warning index that maximizes the utility of thebank is established in this paper, which is the best expecteddefault probability, and correspondingly determines a warn-ing expected default probability interval. For changes in theprobability of default within the warning interval, i.e.,changes in credit status, it is recommended that banks adoptdifferent strategies to reduce losses.

5.5.4. Disciplinary Modules. For lenders at risk levels of loss,the probability of default is high and almost no repaymentis possible. It is recommended that banks determine theprobability of punishment and the amount of punishmentfor such lenders so that their default losses are much less thantheir default gains, but they cannot exceed the level of possi-ble tolerance. At the same time, reasonable intervals forbreach of contract and penalizing the classification ofdefaults by lenders in different situations are recommended,

which will also be recorded in the disciplinary file to increasethe cost of refinancing it to the financial system.

The specific operation and implementation of the abovefunctions can help our dynamic credit tracking model beput in place by constructing a perfect and strict debt collec-tion system, and the detailed steps are completed in turn:the establishment of a personal consumer credit collectiondepartment within the bank, the main function is to trackand collect debts for individuals who do not repay their per-sonal consumer credit in a timely manner. At the same time,it is necessary to strengthen the construction of the network,strengthen the internal information exchange between bankson the credit status of lenders through the interconnection ofthe Internet and information resources, and strengthen therelationship between the bank and the lender’s unit. Thesteps of the debt collection system can be planned like this:

(1) For lenders who do not contact the bank on time orcontact the bank, the bank’s debt collection depart-ment should immediately contact and negotiate withthe lender’s unit. At this time, the bank and the unitcan urge the repayment to assist the tracking systemto play a tracking and urging role

(2) Once the unit is unknown or the lender does notknow where to go, the bank debt collection depart-ment should immediately contact the lender’s family,and the parents provide the lender’s whereabouts andurge the lender to repay the loan, if necessary, by theirparents to return it. The bank debt collection depart-ment should immediately contact the public securitydepartment through a loan lender whose family isstill unable to contact and conduct inquiries nation-wide through its unique identity card number. Inshort, it is necessary to achieve the supervision andcontact functions through the dual channels of familyand society and increase the intensity of tracking

(3) If the above situation still does not work, the bankcan immediately freeze or stop its basic account,recover the loan, and resort to the law if necessary.In the event that a credit loan cannot be recovered,the banking institution will suffer huge losses. Inorder to minimize such losses, the borrower may berequired to provide mortgage guarantees or use creditinsurance and personal credit insurance to pass onsome risks. In addition, personal credit risk can bepassed on through marketization or correspondinginsurance business. These are the risk warning func-tions of the debt collection system. Once the riskwarning signal appears, not only should the above-mentioned functions such as tracking, supervision,and contact be linked, but it should also be trans-ferred to the fourth step, that is, disciplinary actionwill be implemented immediately

(4) The realization of the disciplinary function shouldalso establish a matching reward and punishmentmechanism to assist the dynamic tracking model toimplement debt collection

0 1 2 3 4 5 6 7 8Sample number

0

0.5

1

1.5

2

2.5

3

3.5

4

4.5

Cred

it ra

ting

scor

e

Momentum gradient algorithm simulation result

Simulation dataRaw data

Figure 13: Momentum gradient test results.

13Complexity

Table5:Summarytableof

multiplealgorithm

testresults.

Samplenu

mber

Creditrating

score(original)

LMsimulation

Epo

ch=5

Goal=

1e-6

Simulationeffect

MSE

=2.409e009

Bayesiansimulation

Epo

ch=5

Goal=

1e-6

Simulationeffect

MSE

=3.0933e008

Dynam

icgradient

simulation

Epo

ch=43,381

Goal=

e-2

Simulationeffect

MSE

=0.0100

001

22.00

100.00%

2.00

100.00%

1.97

99.00%

002

23.02

90.00%

2.30

90.00%

2.01

99.67%

003

21.66

88.67%

1.91

97.00%

1.93

97.67%

004

43.88

96.00%

3.48

82.67%

3.61

87.00%

005

22.00

100.00%

2.00

100.00%

1.98

99.33%

006

22.00

100.00%

2.00

100.00%

2.03

99.00%

007

21.81

93.67%

1.91

97.00%

2.05

98.33%

008

11.19

93.67%

1.39

87.00%

1.47

84.33%

Ann

otation:(1)Simulationeffect=

1−|Sim

ulationvalue−Truevalue|÷(m

ax(Truevalue)−min(Truevalue))=

1−|Sim

ulationvalue−Truevalue|÷3.(2)Epo

chs:networktraining

times;goal:networktraining

targeterror;MSE

:error

vector.

14 Complexity

On the one hand, an incentive mechanism with incen-tives is established, which is contrary to the penalty mecha-nism of default, and the purpose is to contrast withpunishment, that is, to form an incentive compatibilitymechanism, rewarding those who default on the lender,rewarding those in order to maximize the effectiveness ofbank credit tracking; the bank should determine the rewardratio and the amount of reward in the credit tracking modelto achieve zero-sum between the bank and the lender.Game equilibrium state: the specific design can be as fol-lows: the preferential treatment for the loan repayment ofthe loaner who pays in advance. Regularly publish onlineor media information on the credit rating of the lender.For lenders with high credit ratings, the bank will reducethe interest rate or reduce the principal discount to promoteits repayment or early repayment, thereby reducing thebank collection cost. For those who have difficulties inrepaying loans, they should be given a principal or interestreduction or a graceful loan period; for the support of thebranch, the western volunteers, etc., in accordance withnational policies, the principal and interest reduction orgrace period should be granted.

On the other hand, a penalty mechanism with noticeand coercion is enacted to force the arrears to repay. Spe-cific measures can be taken as follows: First, lenders whosecredit rating is lower than the warning line are announcedin time in the financial system and warned that the costof refinancing will increase or it will be difficult to finance.Second, for the untrustworthy lender, the breach of contractwill be recorded in the personal credit file of the loan andfurther incorporated into the national personal credit infor-mation system, so that the individual transaction behaviorwill be affected in the future, and the social supervisioneffect of a life-long loss affecting life will be achieved. Third,if the circumstances are serious, the legal liability of thedefaulting borrower will be investigated according to thelaw, and the name of the introducer or witness who failsto perform the duties will be announced. The fourth is toimplement a personal bankruptcy system and a credit guar-antee mechanism. With restrictions on consumption andharsh constraints on bankrupt individuals, people can’ttravel abroad, can’t use credit cards, can’t enjoy loan ser-vices, and can’t buy high-end goods. At the same time, asocial credit guarantee company is established to guaranteepersonal credit.

6. Conclusions

This paper analyzes the status quo of the credit industry anduses mathematical models to demonstrate the inevitability offinancial information industry development of big data. Onthis basis, a dynamic credit tracking model of BP neural net-work based on adaptive genetic algorithm was constructed.Using MATLB software to simulate, it can be seen from thetraining results that the LM training algorithm and Bayesianalgorithm can converge within 10e-6 quickly. While momen-tum gradient algorithm convergence is slow, the overalltraining effect is good, and the simulation results are all up

to 90%, indicating that the model can be well applied to thebig data in Internet finance credit.

Data Availability

The data used to support the findings of this study areavailable from the corresponding author upon request.

Conflicts of Interest

The author declares that there is no commercial or associativeinterest that represents a conflict of interest in connectionwith the work submitted.

Acknowledgments

This research is supported by the Shandong Province NaturalScience Fund Project “the Simulation Research on Industry,Population, Employment, Education and Social SecurityPolicy under the new normal of Shandong province” (Projectnumber ZR2015GL013), the Shandong Social Science Pro-gram Research Project “Study on the Value Orientation andthe United Front Strategy of the Young Generation of PrivateEntrepreneurs of Shandong” (Project number 17CTZJ05),and the Shandong Institute of Business and Technology’steaching reform research project, “Collaborative Educationmode of Applied Talents in Local Universities”—TakingHuman Resources Management as an Example (Projectnumber 11688G201720).

References

[1] R. H. Wang, “Research on personal credit consumer financebased on mobile terminal,” Chinese Commercial Theory,vol. 34, pp. 29-30, 2017.

[2] X. G. Chen and Y. Q. Li, “Research on the transmissionmechanism of internet credit risk based on the perspective ofinternet finance,” Innovation, vol. 11, no. 1, pp. 80–90, 2017.

[3] C. H. Liu, J. Jiang, and J. Li, “Business strategy and perceivedbenefits of internet banking: their impact on banks’ strategicresponses to China’s entry to WTO,” Journal of Dong HuaUniversity, vol. 20, no. 3, pp. 63–68, 2003.

[4] Y. C. Zhang and J. Z. Yang, “Research on internet financialcredit issues in the background of big data,” E-commerce,vol. 5, pp. 55-56, 2016.

[5] W. F. Feng and C. M. Li, “Research on big data credit construc-tion under the background of internet finance,” InternationalFinance, vol. 10, pp. 61–66, 2015.

[6] Z. Chen, “Practice of comprehensive credit reporting and dif-ferentiated development at home and abroad,” Credit Report-ing, vol. 36, no. 06, pp. 50–53, 2018.

[7] G. S. Pan, “Building a well-developed Chinese credit market,”Credit Information, vol. 32, no. 11, pp. 1–4, 2014.

[8] N. Ghatasheh, “Business analytics using random forest treesfor credit risk prediction: a comparison study,” InternationalJournal of Advanced Science and Technology, vol. 72,pp. 19–30, 2014.

[9] R. A. Bartlett, Practitioner’s Guide to Business Analytics: UsingData Analysis Tools to Improve Your Organization’s DecisionMaking and Strategy, McGraw Hill Professional, 2013.

15Complexity

[10] W. Guo, Y. Z. Wang, and J. Y. Qin, “Research on the construc-tion of personal credit reporting system in China under thebackground of big data,” Modern Management Science,vol. 3, pp. 3–8, 2018.

[11] Z. F. Chang, Research on the Construction and Risk Manage-ment of Internet Financial Credit System, Nanjing University,2018.

[12] L. Chen, T. Y. Zhou, J. Y. Ren, and Y. X. Gu, “Research on theapplication of large data credit reporting in the internet back-ground: taking “ant golden clothes” as an example,” MarketWeekly, vol. 7, pp. 108–110, 2018.

[13] J. L. Wen and L. H. He, “Cooperation of credit investigationbetween mainland China and Taiwan under the new situa-tion,” Asian Agricultural Research, vol. 8, no. 10, pp. 9–12,2016.

[14] B. A. Kusi, E. K. Agbloyor, K. Ansah-Adu, and A. Gyeke-Dako,“Bank credit risk and credit information sharing in Africa:does credit information sharing institutions and contextmatter?,” Research in International Business and Finance,vol. 42, pp. 1123–1136, 2017.

[15] B. J. Yang and X. H. Liu, “Analysis of basic credit data mining,”Modern Management Science, vol. 8, pp. 54–56, 2015.

[16] W. Wei, Z. Sun, H. Song, H. Wang, X. Fan, and X. Chen,“Energy balance-based steerable arguments coverage methodin WSNs,” IEEE Access, vol. 6, pp. 33766–33773, 2018.

[17] L. Ji, H. Chi, and J. M. Chen, Application of MultivariableAnalysis in Performance Evaluation of Bank and SavingsDeposits, China Management Science, 2001.

[18] J. Yi, Q. Wang, D. Zhao, and J. T. Wen, “BP neural networkprediction-based variable-period sampling approach for net-worked control systems,” Applied Mathematics and Computa-tion, vol. 185, no. 2, pp. 976–988, 2007.

[19] Y. Wang, T. Q. Luan, and M. Z. Pu, “Optimization of aceticacid fermentation medium based on neural network andgenetic algorithm,” Chinese Journal of Food Science, vol. 12,no. 5, pp. 88–94, 2012.

[20] K. Cui and X. Qin, “Virtual reality research of the dynamiccharacteristics of soft soil under metro vibration loads basedon BP neural networks,” Neural Computing and Applications,vol. 29, no. 5, pp. 1233–1242, 2018.

16 Complexity

Hindawiwww.hindawi.com Volume 2018

MathematicsJournal of

Hindawiwww.hindawi.com Volume 2018

Mathematical Problems in Engineering

Applied MathematicsJournal of

Hindawiwww.hindawi.com Volume 2018

Probability and StatisticsHindawiwww.hindawi.com Volume 2018

Journal of

Hindawiwww.hindawi.com Volume 2018

Mathematical PhysicsAdvances in

Complex AnalysisJournal of

Hindawiwww.hindawi.com Volume 2018

OptimizationJournal of

Hindawiwww.hindawi.com Volume 2018

Hindawiwww.hindawi.com Volume 2018

Engineering Mathematics

International Journal of

Hindawiwww.hindawi.com Volume 2018

Operations ResearchAdvances in

Journal of

Hindawiwww.hindawi.com Volume 2018

Function SpacesAbstract and Applied AnalysisHindawiwww.hindawi.com Volume 2018

International Journal of Mathematics and Mathematical Sciences

Hindawiwww.hindawi.com Volume 2018

Hindawi Publishing Corporation http://www.hindawi.com Volume 2013Hindawiwww.hindawi.com

The Scientific World Journal

Volume 2018

Hindawiwww.hindawi.com Volume 2018Volume 2018

Numerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisAdvances inAdvances in Discrete Dynamics in

Nature and SocietyHindawiwww.hindawi.com Volume 2018

Hindawiwww.hindawi.com

Di�erential EquationsInternational Journal of

Volume 2018

Hindawiwww.hindawi.com Volume 2018

Decision SciencesAdvances in

Hindawiwww.hindawi.com Volume 2018

AnalysisInternational Journal of

Hindawiwww.hindawi.com Volume 2018

Stochastic AnalysisInternational Journal of

Submit your manuscripts atwww.hindawi.com