CN106779755A - A kind of network electric business borrows or lends money methods of risk assessment and model - Google Patents

Abstract

The invention discloses a kind of network electric business debt-credit methods of risk assessment and model, method is comprised the following steps：The historical trading data of collection network electric business debt-credit client is used as sample set；The twin support vector cassification model of least square is built, and it is trained, set up network electric business debt-credit Credit Model；Risk model is borrowed or lent money according to the network electric business set up, forecast sample concentrates the promise breaking label of each sample；For each sample, will predict that the promise breaking label of the promise breaking label and reality for obtaining is compared, determine the predicated error of sample, and the size of each sample weights is determined according to the size of predicated error；Based on each sample weights, the twin support vector cassification model of weighted least-squares is built, re-establish network electric business debt-credit risk model, debt-credit risk is estimated.The present invention assessment degree of accuracy is high.

Description

A kind of network electric business borrows or lends money methods of risk assessment and model

Technical field

The present invention relates to a kind of network electric business debt-credit methods of risk assessment and model.

Background technology

Internet finance is a kind of brand-new Finance Service, and traditional small amount debt-credit service is transferred to internet by it , it is necessary to the crowd for borrowing money can be lent ability and be ready to be lent based on certain condition in network loan platform searching on platform Crowd, due to network loan have it is efficient, simple to operate, towards middle low layer take in crowd (for borrower) and peace The advantage of entirely, transparent, income (for lender) high, thus network loan platform once release just get the nod rapidly with Development^[1].But compared with traditional pattern of lending, network loan there is also shortcoming, such as network loan is unsecured loan, Lender's major part is all general population, the Investment ＆ Financing knowledge without specialty, and some borrowers are in order at the mesh swindled Carrying out provide a loan, this just to lender bring network loan risk, i.e. borrower regulation cut-off refund the date in, due to A variety of causes cannot on time be paid off a loan according to lending agreement, and the possibility of monetary losses is brought to lender^[2]。

Recent years, many scholars propose many new assessing credit risks moulds for the debt-credit risk control of network electric business Type.These models can generally be divided into two classes：The first kind is application level analytic approach^[5], Field Using Fuzzy Comprehensive Assessment^[6]Deng statistics The credit indicator evaluation system that method is set up, and specify the subjective weight of each index.Equations of The Second Kind is to use linear data There are some general character in the risk evaluation model that mining algorithm is set up, these models：Assessment accuracy rate is low, and tracing it to its cause is There is nonlinear organization in credit data.And the two sorting algorithm applications that can process nonlinear organization data more widely have： SVMs^[7]And BP neural network^[8].But be present the unbalanced feature of class in network electric business loan credit data, such as clap Bat loan platform is normally refunded to record and is about 10 times that promise breaking is recorded, for such risk data, supporting vector machine model Error type I False Rate is higher, i.e., normal refund client is judged to the client that breaks a contract.BP neural network there is also same asking Topic, and BP neural network model is when only having the error of training set and forecast set close, and model is just with generalization ability.

Therefore, it is necessary to design a kind of accuracy rate network electric business debt-credit Risk Forecast Method and model higher.

The content of the invention

Technical problem solved by the invention is, in view of the shortcomings of the prior art, there is provided a kind network electric business borrows or lends money wind Dangerous appraisal procedure and model, predictablity rate are high.

Technical scheme provided by the present invention is：

A kind of network electric business borrows or lends money methods of risk assessment, comprises the following steps：

Step 1, the data of collection history debt-credit client are used as sample set；The data of each debt-credit client include that n is tieed up and evaluate Achievement data and promise breaking label target, target=1 represent loan defaults, and target=0 represents normal refund；

Step 2, the data in sample set are pre-processed；The pretreatment includes that missing values treatment, nonnumeric type are commented Valency achievement data quantifies and data normalization；

Step 3, the sample in sample set is divided into according to the value of promise breaking label target by two classes, the first kind is target= 0 normal refund client, matrix is built by its corresponding evaluation index dataEquations of The Second Kind is target=1 Loan defaults client, build matrix by its corresponding evaluation index dataWherein m1 and m2 are respectively two Sample size in class sample,It is n-dimensional vector, represents i-th evaluation index of sample in kth class；

Step 4, based on two class samples, build the twin support vector cassification model of least square, and it is trained, Set up network electric business debt-credit risk evaluation model；

Step 5, the network electric business debt-credit risk evaluation models according to 4 foundation, forecast sample concentrate the promise breaking of each sample Label target；

Step 6, for each sample, it is actual separated that the promise breaking label target and step 1 obtained according to 5 predictions are collected About label target determines the predicated error of sample, and the size of each sample weights is determined according to the size of predicated error；Sample The determination principle of this weight is：The larger sample weights of predicated error are smaller, and the less sample weights of predicated error are larger；

Step 7, based on each sample weights, build the twin support vector cassification model of weighted least-squares, build again Vertical network electric business debt-credit risk evaluation model；

Step 8, the n dimension evaluation index data for gathering new debt-credit client, substitute into the network electricity that step 7 determines after pretreatment Loan risk evaluation model is negotiated to borrow, the corresponding promise breaking label target of borrower is connected to, predicts whether debt-credit client can provide a loan separated About, it is estimated with to this debt-credit risk.

The step 4 specifically includes following steps：

Step 4.1, the structure twin support vector cassification model of least square：

s.t.-(K(B,C^T)w₁+e₂b₁)=e₂-η₂

s.t.K(A,C^T)w₂+e₁b₂=e₁-η₁

Wherein, C₁And C₂It is punishment parameter, optimal value is selected (i.e. first randomly by according to ten folding cross validations It is 10 mutually disjoint size identical subsets to data cutting, then using 9 data training patterns of subset (training mould C in type₁And C₂Span be：C₁∈[2^-5,…,2⁵], C₂∈[2^-5,…,2⁵]), tested using remaining 1 subset Model；This process is repeated to possible 10 kinds of selections, average test error in 10 evaluation and tests is finally selected minimum Model)； η₁And η₂It is loss variable (being also called error)；e₁And e₂1 column vector is all for element；K(X, C^T) it is gaussian kernel function, X=A or B (σ is the width parameter of function, σ>0, σ according to Ten folding cross validations select optimal value)；w₁And w₂It is weight vector, b₁And b₂Presentation class threshold value；w₁、w₂、b₁And b₂To wait to ask Parameter；

Step 4.2, w is asked for by below equation₁、w₂、b₁And b₂：

Wherein, U=[K (A, C^T),e₁], V=[K (B, C^T),e₂]；

Step 4.3, according to fixed parameter set up with lower network electric business borrow or lend money risk model：

Wherein, x represents the n dimension evaluation index data of pretreated new reception client, and target (x)=0 represents pre- Surveying the new debt-credit client can normally refund, and target (x)=1 represents predicts that the new debt-credit client understands loan defaults.

In the step 6, the computing formula of sample weights is：

Wherein, | η_ki| (k=1,2) is i-th predicated error of sample in sample set kth class sample, and δ is | η_ki|(k =standard deviation 1,2).

The step 7 specifically includes following steps：

Step 7.1, the structure twin support vector cassification model of weighted least-squares：

The twin support vector cassification model (WLS-TSVM1) of first weighted least-squares:

s.t.-(K(B,C^T)w'₁+e₂b₁')=e₂-η₂

The twin support vector cassification model (WLS-TSVM2) of second weighted least-squares:

s.t.K(A,C^T)w'₂+e₁b'₂=e₁-η₁

Wherein, C₁' and C'₂It is punishment parameter, optimal value is selected according to ten folding cross validations；w'₁And w'₂For weigh to Amount, b'₁And b'₂Presentation class threshold value；w'₁、w'₂、b'₁And b'₂It is parameter to be asked；

Step 7.1, w' is asked for by below equation₁、w'₂、b'₁、b'₂：

Wherein, M=[K (A, C^T),e₁], N=[K (B, C^T),e₂]；

P₁ ^-1WithIt is respectively with ρ_1iAnd ρ_2iIt is the diagonal matrix of the elements in a main diagonal；

Step 7.1, according to fixed parameter re-establish with lower network electric business borrow or lend money risk model：

Wherein, K (x^T,C^T) it is gaussian kernel function；σ is the width parameter of function, σ>0, σ selects optimal value according to ten folding cross validations；X represents that the n of debt-credit client new after pre-processing ties up evaluation index number According to target'(x)=0 represent and predict that the new debt-credit client can normally refund, target'(x)=1 represent prediction this is new Debt-credit client may loan defaults.

A kind of network electric business borrows or lends money risk evaluation model, it is characterised in that the model is:

Wherein, x represents the n dimension evaluation index data of pretreated new debt-credit client, target'(x)=0 expression is in advance Surveying the new debt-credit client can normally refund, and this debt-credit risk is relatively low；Target'(x debt-credit visitor for predicting that this is new)=1 is represented Family meeting loan defaults, this debt-credit risk is higher；Model parameter is solved by above step.

Principle of the invention is illustrated below：

The twin support vector cassification model of 1 least square

Twin SVMs (twins support vector machine, be abbreviated as TSVM) is in SVMs (SVM) put forward on the basis of, the training speed its purpose is to improve SVM.The basic thought of TSVM is to positive and negative class Training sample constructs two nonparallel Optimal Separating Hyperplanes, and each Optimal Separating Hyperplane is meeting the condition away from another kind of data point Under be fitted this class data point as far as possible^[9].Compared with SVM, there are two advantages in TSVM：First, it is equal in positive and negative class number of samples In the case of, the time complexity of SVM is O (m³), the time complexity of TSVM is O (2* (m/2)³), wherein m is training set sample This number, from time complexity it can be seen that the training speed of TSVM improves 4 times compared to SVM^[10]；Second, it is uneven for class The data of weighing apparatus, the influence that corresponding punishment parameter elimination class imbalance is brought is set when hyperplane is sought to each hyperplane^[11]。 Due to the classification advantage of the two uniquenesses of TSVM, TSVM is widely applied to Image detection in recent years^[12], chemical analysis^[13]And well Lower diagnosis^[14]Etc. industry.

The training sample of given m n dimensions, is classified as two classes, and the wherein first kind is designated as matrixThe Two class samples are designated as matrixWherein m1 and m2 are respectively the sample size in two class samples,For n tie up to Amount, represents j-th sample of the i-th class；Then TSVM models are described in detail below：

In above formula, C₁And C₂It is punishment parameter, optimal value is selected according to ten folding cross validations；η₁And η₂For loss becomes The difference of amount (being also called error), i.e. predicted value and actual value；e₁,e₂1 column vector is all for element；C=[A^T,B^T]^T, K (X, C^T) it is gaussian kernel function, X=A or B；σ is the width parameter of function, σ>0, σ according to Ten folding cross validations select optimal value；w₁And w₂It is weight vector, b₁And b₂Presentation class threshold value, w₁、w₂, b₁And b₂To wait to ask Parameter.

The twin SVMs of least square (Least squares twins support vector machine, letter It is written as LS-TSVM) the least square thought is introduced, the slack variable η in LS-TSVM₁And η₂C has been used respectively₁/ 2, C₂The 2 of/2 Normal form, can thus omit η₁>=0 and η₂>=0 constraints, while inequality constraints is revised as equality constraint, so Quadratic programming problem conversion is just reduced the computation complexity of twin SVMs for Solving Linear problem.

Here is the specific descriptions to non-linear LS-SVM models：

s.t.-(K(B,C^T)w₁+e₂b₁)=e₂-η₂ (4)

s.t.K(A,C^T)w₂+e₁b₂=e₁-η₁ (6)

(4) formula and (6) formula are substituted into (3) formula respectively and (5) formula is obtained：

(7) formula is asked on w respectively₁And b₁Derivative, and make it be equal to zero, draw

K(A,C^T)^T[K(A,C^T)w₁+e₁b₁]+C₁K(B,C^T)^T[K(B,C^T)w₁+e₂b₁+e₂]=0 (9)

It is matrix form to merge (9) formula and (10) formula, is obtained

Can be solved by (11) formula

In above formula：

U=[K (A, C^T),e₁]

V=[K (B, C^T),e₂]

Can similarly obtain：

(12) formula and (13) formula are substituted into following (14) formula respectively, two Optimal Separating Hyperplanes just can be obtained.For new Sample number strong point, according to sample number strong point to two range estimation data point generics of Optimal Separating Hyperplane, data point from Which hyperplane is nearer, and which kind of the data point just belongs to.

K(x^T,C^T)w₁+b₁=0, K (x^T,C^T)w₂+b₂=0 (14)

The twin support vector cassification model of 2 weighted least-squares

The twin SVMs of 2.1 weighted least-squares

Although LS-TSVM algorithms solve the problems, such as that TSVM calculates complicated, but when training set has data exception point, The punishment parameter of formed objects is assigned to each sample in similar, algorithm robustness is reduced.In order to reduce data exception point Influence to Optimal Separating Hyperplane, proposes that a kind of twin SVMs of weighted least-squares is calculated on the basis of LS-TSVM herein Method, improves the not good enough problem of LS-TSVM algorithm robustness.Weighting treatment has been made to error in (3) formula and (5) formula, if first Class sample and the corresponding weights of Equations of The Second Kind sample are respectively ρ_1iAnd ρ_2i, then corresponding optimization problem can be described as

WLS-TSVM1:

WLS-TSVM2:

The Lagrangian Form of WLS-TSVM1 formulas can be expressed as：

Wherein, α is m₂The Lagrange multiplier of dimension.

Obtained according to KTT conditions：

Obtained by (18) formula and (19) formula：

Order

M=[K (A, C^T),e₁] (23)

N=[K (B, C^T),e₂] (24)

Can then be drawn by (22) formula：

By (20) formula, (21) Shi Ke get：

Wherein, P₁ ^-1It is with (ρ₁)_iiIt is the diagonal matrix of the elements in a main diagonal.

M is introduced to WLS-TSVM2 formulas₁The β Lagrange multipliers of dimension, can similarly obtain：

Wherein P₂ ^-1It is with (ρ₂)_iiIt is the diagonal matrix of the elements in a main diagonal.

S is determined according to (25) formula and (27) formula₁And s₂, thus just two hyperplane can be determined by (14) formula.Weighting is most The classification function that a young waiter in a wineshop or an inn multiplies twin SVMs can be expressed as：

Above calculating process establishes the twin vector machine Credit Risk Model of weighted least-squares, for new client, Then collection customer information calculates the value for determining target, target'(x as the x vector inputs in (29) formula)=0 representative Sample belongs to the first kind, that is, predict that the client can normally refund；Target'(x)=1 representative sample Equations of The Second Kind, the i.e. client May loan defaults.

2.2 weights are set

In order to reduce influence of the data exception point to disaggregated model so as to improve the generalization ability of disaggregated model, while again not Ignore effect of the data exception point to disaggregated model, herein using normpdf be more than zero the characteristics of, according to The larger sample weights of probability density function values calculating sample weights, i.e. predicated error are smaller, the less sample power of predicated error Value is larger, and influence of the data exception point to disaggregated model is reduced with this.

Specific weights set as follows：

Wherein, | η_ki| (k=1,2) is i-th predicated error of sample in sample set kth class sample, and δ is | η_ki|(k =standard deviation 1,2).

Specific weights solution procedure is as follows：

The first step：The twin support vector cassification model of least square is set up using training set data；

Second step：According to the first step set up the twin support vector cassification model of least square to training set data again Classification, obtains classification results；

3rd step：The classification results obtained according to the actual classification of training set sample and second step calculate training set sample and miss Difference；

4th step：Sample weights are calculated according to (30) formula.

The 3 network electric business debt-credit Credit Risk Assessment Models based on WLS-TSVM

Network electric business debt-credit Credit Risk Model estimation flow is as follows：Transaction data is borrowed or lent money according to existing electric business history first Built one's credit risk evaluation model, and the credit standing of borrower is then assessed with Credit Risk Model.By to having trained Data set sets up disaggregated model, and new borrower is divided into two classes by application model：Default risk is high and default risk is low, for Default risk borrower's refusal loaning bill project high passes through, so as to reduce default risk.

The specific execution step of the network electric business debt-credit Credit Risk Model based on WLS-TSVM algorithms is as follows：

(data of history debt-credit client can be from existing network electric business loan platform for the data of collection history debt-credit client Obtain) used as training sample, the data to be collected include that (target=1 is represented for evaluation index data and promise breaking label target Loan defaults, target=0 represents normal refund), evaluation index data are specifically included：

Borrower's essential information field (name, sex, identification card number, cell-phone number, No. QQ, conventional mailbox, marital status, Credit number, Employment, schooling, contact person etc.)；

Network behavior field (accesses type, access times, average access duration of website etc.)；

Academic student status field (academic type, academic level, graduation universities and colleges, graduation department, graduation time, major name, Habit form, length of schooling etc.)；

(borrower is in third party's financial platform history loaning bill number of times, history loaning bill success rate, credit for third-party platform field Grade etc.)；

Social networks field (friend quantity, friend species number, friend averaged historical of the borrower in social platforms such as Tengxuns Loaning bill number of times, friend's history loaning bill success rate, friend's average credit grade scoring etc.)；

Loaning bill information field (borrowing balance, the life of loan, borrowing rate, use of the loan, credit grade, be in debt income ratio, It is engaged in the time of work, house now and possesses situation, annual income, history loaning bill promise breaking label etc.)；

2. the treatment of data prediction, including missing values, the digitlization of nonnumeric field, data normalization：

2.1. missing values treatment, missing values are deleted more than the row of training set data total amount 60%；Missing values are less than instruction Practice the row of collection data total amount 60%, missing values use -1 is filled up；

2.2. nonnumeric field digitlization：

Time map is year, month, day；

Place name is mapped as longitude and latitude according to national longitude and latitude table；

Ordinal data is mapped as ordinal number, and (such as schooling can be divided into illiteracy, primary school, junior middle school, senior middle school, university etc., text Blind->0, primary school->1, junior middle school->2, senior middle school->3, university->4)；

Nominal level variable is mapped as 0-1 dummy variables, and (it is 1 that such as sex sets female, and man is for 0)；

2.3. it is standardized with zscore function pair data, eliminates the isomerism between data；

3. training set data is divided into two classes by the value according to promise breaking label target, and the first kind is normal for target=0's Refund client, is represented with matrix A, and Equations of The Second Kind is the loan defaults client of target=1, is represented with matrix B；Pre- by data Value of the training set data for the treatment of in addition to target according to target constitutes a line of matrix A or B, i.e., It is n-dimensional vector, represents i-th evaluation index of sample in kth class；

4. application training collection data, according to the twin support vector cassification model of least square, set up the debt-credit of network electric business Credit Risk Model：

3.1. C matrixes, C=[A are calculated^T,B^T]^T；

3.2. K (A, C are determined according to gaussian kernel function^T) and K (B, C^T)；

3.3. calculating matrix U and V；

3.4. by the C of ten folding cross validation selection sort errors minimum₁And C₂, and determined according to (12) formula and (13) formula Final w and the value of b；

3.5. two Optimal Separating Hyperplanes are determined according to (14)；

5. according to the promise breaking label of the 4 twin support vector cassification model prediction training sets of least square set up, i.e. root The value of the target of electric business client is determined according to (29) formula；

6. the target values according to the 5 target values for obtaining and training set electric business client's real trade determine sample error, And sample weights ρ is determined according to (30) formula_1i、ρ_2i、P₁ ^-1And P₂ ^-1；

7. application training collection data re-establish network electricity according to the twin support vector cassification model of weighted least-squares Negotiate to borrow loan Credit Risk Model：

7.1. calculating matrix M and N；

7.2. Lagrange multiplier α and β are determined according to (30) formula and (32) formula；

7.3. by the C of ten folding cross validation selection sort errors minimum₁' and C₂', and it is true according to (29) formula and (31) formula The value of fixed final w and b；

7.4. two Optimal Separating Hyperplanes are determined according to (14)；

8. the value of the target of new electric business client, target'(x are determined according to (29) formula)=0 represent and predict the visitor Whether family can normally refund, target'(x)=1 represent predict the client may loan defaults.

Beneficial effect：

The present invention proposes a kind of twin based on weighted least-squares on the basis of the twin SVMs of least square (Weighted least squares twins support vector machine, are abbreviated as WLS- to SVMs TSVM network electric business debt-credit risk evaluation model).The twin supporting vector machine model of weighted least-squares takes full advantage of twin The advantage that SVMs can quick and precisely classify to the unbalanced risk data of class, introduces the least square thought and simplifies Calculating during twin model construction of SVM, while having done weighting treatment to the error in twin SVMs, reduces Influence of the data exception point to Optimal Separating Hyperplane, so as to improve nicety of grading, largely improves risk profile accurate True rate.Shown by the experiment of actual data analysis, WLS-TSVM algorithms reduce shadow of the noise data to disaggregated model really Ring, model training speed and classification accuracy are improved.

Brief description of the drawings

Fig. 1 is the roc curve maps of BP neural network；

Fig. 2 is the roc curve maps of SVM；

The roc curve maps of Fig. 3 WLS-TSVM of the invention.

Specific embodiment

The present invention is described in more detail below in conjunction with the drawings and specific embodiments.

In the present embodiment, from the credit risk data for patting network loan industry in loan " witch mirror cup " air control algorithm contest As experimental data.

(1) data set is constituted

Value according to the label target that broken a contract in experimental data may determine that the financial status of borrower, target=1 generations Table loan defaults, target=0 represents normal refund.Debt-credit data record is randomly selected from experiment competition data to respectively constitute The training set and test set of model, data set have six big field classifications, data set after each big field type subdivision Totally 207 data dimensions, training set and test set data distribution are as shown in table 1.

The data set of table 1 is constituted

(2) data prediction

Due to there is character type data in risk data, and matlab is unable to the data of processing character type, so right herein The numeral of character type is processed, and the address field of character type has been substituted for into the corresponding longitude and latitude in address.

The missing values number of each row of risk data table is counted with histc functions, 18000 have been more than for missing values number The row of (data total amount 60%), are deleted, by treatment, 174 attributes of final residue.

It is standardized from zscore function pair data in testing herein, the difference between heterogeneity data is eliminated with this It is different.

(3) interpretation

Experimental situation:Win7 operating systems, 6G internal memories, CPU2.5Hz, matlab2014a.To add with matlab instruments The power twin algorithm of support vector machine of least square (WLS-TSVM), SVMs (SVM) and BP neural network are on forecast set Prediction classification accuracy compare, the performance of model is measured according to following two indexs：(1) classification accuracy；(2) The model training time.Table 2 is three kinds of time complexity contrasts of algorithm, and wherein m is sample total number, and m1, m2 are respectively positive and negative Class number of samples, n is data dimension.Table 3 is the Comparative result of the prediction classification of SVM and WLS-TSVM on data set.

The Algorithms T-cbmplexity of table 2 is contrasted

The experimental result of table 3 is contrasted

From table 3 it can be seen that in terms of predictablity rate, the predictablity rate of WLS-TSVM models is higher than SVM and BP god Through the predictablity rate of network.Time complexity analysis contrast according to table 2, it can be seen that the time complexity of WLS-TSVM is small Take in the execution time major part of SVM and the time complexity of BP neural network, and algorithm and spend in the training of model.Pass through Experimental data in table 3, we are really it can be seen that in terms of training time length, the WLS-TSVM training times are less than really The training time of SVM and BP neural network model.

Fig. 1~Fig. 3 is respectively the roc curve maps and WLS-TSVM of the invention of the roc curve maps of BP neural network, SVM Roc curve maps；The abscissa of roc curves is FPR (false positive rate, negative and positive class rate), and ordinate is TPR (true positive rate, real class rate).With roc (Receiver Operating Characteristic) curves and Auc values can be with the overall classification performance of classification device, and from Fig. 1, Fig. 2, the roc curve maps in Fig. 3 can be seen that WLS- Auc ((Area Under roc Curve)) value (area under auc values correspondence blue curve) of TSVM is neural apparently higher than BP The auc values of network and SVM algorithm.

Network electric business debt-credit Credit Risk Assessment Model knot of the present invention based on the twin SVMs of weighted least-squares Having closed twin SVMs can process the advantage of the unbalanced nonlinear organization data of class, with weighted least-squares method letter Change the calculating of twin SVMs, while having done weighting treatment to error term, reduce data exception point to disaggregated model Influence, so not only increase classification predictablity rate, and reduce the model training time.Experimental section is patting loan As a example by risk data in " witch mirror cup " air control contest, illustrate that the network electric business debt-credit risk model based on WLS-TSVM can be with Satisfied nicety of grading is obtained with the less training time.Illustrate that this model has herein by theory and practice preferable Data fitting effect, be with a wide range of applications.

Bibliography

[1] money gold leaf, current situation and prospect [J] finance forum, 2012 (1) of Yang Fei China P2P network loans: 45-51.

[2]Xinhui Chen,Xiangxiang Ji,Xinrui Zhang.A Contrastive Study of P2P Network Lending in China and American[C].Asia-Pacific E-Business and E- Government Conference(APEE2014),2014.

[3]Tang Guolei.P2P network lending in the credit riskrating of the individual[J].International English Education Research,2015(9):23-26.

[4]Fengzhi Liu,MingyanWang.The credit risk of theborrower’s credit in P2P network[J].International English Education Research,2015(8):72-75.

[5]Niejin,Hulinshuang.P2P network lending platform evaluationbased onlink analysis[J].Research on Library Science,2015(13):37-45.

[6]Liqun You,Yudan Zhu.P2P network lending platform business model evaluation[J].Journal of Commercial Economics,2015(32):86-88.

[7]JinYu,ZhuYudan.Data-Driven Approach to Predict Default Risk of Loan for Online Peer-to-Peer(P2P)Lending[C].Communication Systems and Network Technologies(CSNT),2015Fifth International Conference on,2015.

[8]DunGang Zang,MngYu Qi,YanMing Fu.The credit risk assessment of P2P lending based on BP neural network[C].The 2014International Conference on Industrial Engineering and Management Science(IEMS2014),2014.

[9]Xinjun Peng,A ν-twin support vector machine(ν-TSVM)classifier and its geometric algorithms[J].Information Sciences,2010,180(20):3863-3875.

[10]Xijiong Xie,Shiliang Sun.Multitask centroid twins supportvectormachines[J].Neurocomputing,2015,149:1085-1091.

[11]Yuan-Hai Shao,Xiang-Yu Hua,Li-Ming Liu,Zhi-Min Yang,Nai-Yang Deng.Combined outputs framework for twins supportvectormachines[J] .AppliedIntelligence,2015,43(2):424-438.

[12] Wei little Li, Liu Jinghua, yellow shake green grass or young crops applications Raman spectrum and imaging technique observe α particle irradiations to Rhizopus oryzae Role and influence [J] University of Anhui's journal (natural science edition), 2014,38 (1):66-72.

[13] Qiu Mengchan, Su Chengli, Zhong Guocai, beam Jianping are based on KICA-TSVR method cracking of ethylene product and are distributed soft survey Amount [J] control engineerings, 2015,25 (33):458-464.

[14] Dong Hongbo, Shen Zhongjie, Yao Ya peak are based on underground coal mine rig bearing failure diagnosis [J] the collieries of TWSVM Machinery, 2015,36 (5):298-300.

Claims (5)

1. a kind of network electric business borrows or lends money methods of risk assessment, it is characterised in that comprise the following steps：

Step 1, the data of collection history debt-credit client are used as sample set；The data of each debt-credit client include that n ties up evaluation index Data and promise breaking label target, target=1 represent loan defaults, and target=0 represents normal refund；

Step 2, the data in sample set are pre-processed；The pretreatment includes that missing values treatment, nonnumeric type evaluation refer to Mark data quantify and data standardization；

Step 3, the sample in sample set is divided into according to the value of promise breaking label target by two classes, the first kind is target=0's Normal refund client, matrix is built by its corresponding evaluation index dataEquations of The Second Kind is target=1's Loan defaults client, matrix is built by its corresponding evaluation index dataWherein m1 and m2 are respectively two classes Sample size in sample,It is n-dimensional vector, represents i-th evaluation index of sample in kth class；

Step 4, based on two class samples, build the twin support vector cassification model of least square, and it is trained, set up Network electric business borrows or lends money risk evaluation model；

Step 5, the network electric business debt-credit risk evaluation models according to 4 foundation, forecast sample concentrate the promise breaking label of each sample target；

Step 6, the actual promise breaking mark collected for each sample, the promise breaking label target and step 1 obtained according to 5 predictions Sign target and determine the predicated error of sample, and the size of each sample weights is determined according to the size of predicated error；Sample is weighed Weight determination principle be：The larger sample weights of predicated error are smaller, and the less sample weights of predicated error are larger；

Step 7, based on each sample weights, build the twin support vector cassification model of weighted least-squares, re-establish net Network electric business borrows or lends money risk evaluation model；

Step 8, the n dimension evaluation index data for gathering new debt-credit client, the network electric business that step 7 determination is substituted into after pretreatment are borrowed Borrow risk evaluation model, be connected to the corresponding promise breaking label target of borrower, predict debt-credit client whether can loan defaults, with This debt-credit risk is estimated.

2. network electric business according to claim 1 borrows or lends money methods of risk assessment, it is characterised in that the step 4 is specifically wrapped Include following steps：

Step 4.1, the structure twin support vector cassification model of least square：

$\underset{w_1,b_1}{min}\frac{1}{2}\left|\right|K(A,C^T)w_1+e_1{b}_1|{|}^2+\frac{1}{2}{C}_1{\mathrm{\η}}_2^T{\mathrm{\η}}_2$

s.t.-(K(B,C^T)w₁+e₂b₁)=e₂-η₂

$\underset{w_2,b_2}{min}\frac{1}{2}\left|\right|K(B,C^T)w_2+e_2{b}_2|{|}^2+\frac{1}{2}{C}_2{\mathrm{\η}}_1^T{\mathrm{\η}}_1$

s.t.K(A,C^T)w₂+e₁b₂=e₁-η₁

Wherein, C₁And C₂It is punishment parameter, optimal value is selected according to ten folding cross validations；η₁And η₂It is loss variable；e₁With e₂1 column vector is all for element；C=[A^T,B^T]^T,K(X,C^T) it is gaussian kernel function, X=A or B；w₁And w₂It is weight vector, b₁ And b₂Presentation class threshold value；w₁、w₂、b₁And b₂It is parameter to be asked；

Step 4.2, w is asked for by below equation₁、w₂、b₁And b₂：

$\left[\begin{array}{c}{w}_1\\ b_1\end{array}\right]=-{(V^{T}V+\frac{1}{C_1}{U}^{T}U)}^{-1}{V}^T{e}_2$

$\left[\begin{array}{c}{w}_2\\ b_2\end{array}\right]={(U^{T}U+\frac{1}{C_2}{V}^{T}V)}^{-1}{U}^T{e}_1$

Wherein, U=[K (A, C^T),e₁], V=[K (B, C^T),e₂]；

Step 4.3, according to fixed parameter set up with lower network electric business borrow or lend money risk model：

$t\arg et\left(x\right)=\left\{\begin{array}{ccc}0& if& |K(x^T,C^T)w_1+b_1|<|K(x^T,C^T)w_2+b_2|\\ 1& if& |K(x^T,C^T)w_1+b_1|>|K(x^T,C^T)w_2+b_2|\end{array}\right.;$

Wherein, x represents the n dimension evaluation index data of pretreated new reception client, and target (x)=0 represents that prediction should New debt-credit client can normally refund, and target (x)=1 represents predicts that the new debt-credit client understands loan defaults.

3. network electric business according to claim 2 borrows or lends money methods of risk assessment, it is characterised in that in the step 6, sample The computing formula of weight is：

${\mathrm{\&rho;}}_{ki}=\frac{1}{\sqrt{2\mathrm{\&pi;}}\mathrm{\&sigma;}}\exp \&lsqb;-\frac{|{\mathrm{\&eta;}}_{ki}{|}^2}{2{\mathrm{\&delta;}}^2}\&rsqb;$

Wherein, | η_ki| (k=1,2) is i-th predicated error of sample in sample set kth class sample, and δ is | η_ki| (k=1,2) Standard deviation.

4. network electric business according to claim 3 borrows or lends money methods of risk assessment, it is characterised in that the step 7 is specifically wrapped Include following steps：

Step 7.1, the structure twin support vector cassification model of weighted least-squares：

The twin support vector cassification model (WLS-TSVM1) of first weighted least-squares:

$\underset{w_1,b_1}{min}\frac{1}{2}\left|\right|K(A,C^T)w_1^{\&prime;}+e_1{b}_1^{\&prime;}|{|}^2+\frac{1}{2}{C}_1^{\&prime;}\underset{i=1}{\overset{m_2}{\&Sigma;}}{\mathrm{\&rho;}}_{1i}{\mathrm{\&eta;}}_{2i}^2$

s.t.-(K(B,C^T)w'₁+e₂b′₁)=e₂-η₂

The twin support vector cassification model (WLS-TSVM2) of second weighted least-squares:

$\underset{w_2,b_2}{min}\frac{1}{2}\left|\right|K(B,C^T)w_2^{\&prime;}+e_2{b}_2^{\&prime;}|{|}^2+\frac{1}{2}{C}_2^{\&prime;}\underset{i=1}{\overset{m_1}{\&Sigma;}}{\mathrm{\&rho;}}_{2i}{\mathrm{\&eta;}}_{1i}^2$

s.t.K(A,C^T)w'₂+e₁b'₂=e₁-η₁

Wherein, C '₁And C'₂It is punishment parameter, optimal value is selected according to ten folding cross validations；w′₁And w'₂It is weight vector, b '₁ And b'₂Presentation class threshold value；w′₁、w'₂、b′₁And b'₂It is parameter to be asked；

Step 7.1, w ' is asked for by below equation₁、w'₂、b′₁、b'₂：

$\left[\begin{array}{c}{w}_1^{\&prime;}\\ b_1^{\&prime;}\end{array}\right]=-{\left(M^{T}M\right)}^{-1}{N}^T\mathrm{\&alpha;}$

$\left[\begin{array}{c}{w}_2^{\&prime;}\\ b_2^{\&prime;}\end{array}\right]={\left(N^{T}N\right)}^{-1}{M}^T\mathrm{\&beta;}$

Wherein, M=[K (A, C^T),e₁], N=[K (B, C^T),e₂]；

$\mathrm{\&alpha;}={\&lsqb;\frac{P_1^{-1}}{C_1^{\&prime;}}+N{\left(M^{T}M\right)}^{-1}{N}^T\&rsqb;}^{-1}{e}_2;$

$\mathrm{\&beta;}={\&lsqb;\frac{P_2^{-1}}{C_2^{\&prime;}}+M{\left(N^{T}N\right)}^{-1}{M}^T\&rsqb;}^{-1}{e}_1;$

WithIt is respectively with ρ_1iAnd ρ_2iIt is the diagonal matrix of the elements in a main diagonal；

Step 7.1, according to fixed parameter re-establish with lower network electric business borrow or lend money risk model：

${\mathrm{target}}^{\&prime;}\left(x\right)=\left\{\begin{array}{ccc}0& if& |K(x^T,C^T)w_1^{\&prime;}+b_1^{\&prime;}|<|K(x^T,C^T)w_2^{\&prime;}+b_2^{\&prime;}|\\ 1& if& |K(x^T,C^T)w_1^{\&prime;}+b_1^{\&prime;}|>|K(x^T,C^T)w_2^{\&prime;}+b_2^{\&prime;}|\end{array}\right.$

Wherein, K (x^T,C^T) it is gaussian kernel function；X represents that the n of debt-credit client new after pre-processing ties up evaluation index data, Target'(x)=0 represent predict the new debt-credit client can normally refund, target'(x)=1 represent predict the new debt-credit Client may loan defaults.

5. a kind of network electric business debt-credit risk evaluation model, it is characterised in that the model is:

${\mathrm{target}}^{\&prime;}\left(x\right)=\left\{\begin{array}{ccc}0& if& |K(x^T,C^T)w_1^{\&prime;}+b_1^{\&prime;}|<|K(x^T,C^T)w_2^{\&prime;}+b_2^{\&prime;}|\\ 1& if& |K(x^T,C^T)w_1^{\&prime;}+b_1^{\&prime;}|>|K(x^T,C^T)w_2^{\&prime;}+b_2^{\&prime;}|\end{array}\right.$

Wherein, x represents the n dimension evaluation index data of pretreated new debt-credit client, target'(x)=0 expression prediction should New debt-credit client can normally refund, and this debt-credit risk is relatively low；Target'(x)=1 represent and predict new debt-credit client's meeting Loan defaults, this debt-credit risk is higher；

The step of model parameter is by any one of Claims 1 to 4 1~7 is solved.