CN106097094A - A kind of man-computer cooperation credit evaluation new model towards medium-sized and small enterprises - Google Patents

Abstract

A kind of man-computer cooperation credit evaluation new model towards medium-sized and small enterprises, the topological structure of neutral net is determined according to the dimension in data set, the weights of BP neutral net and threshold value, data input pretreatment have an input layer, one hidden layer and the feedforward neural network of an output layer, any one nonlinear function can be approached with arbitrary accuracy, draw the classification results of PSO BP algorithm, then fuzzy synthesis Weight algorithm is used to calculate the classification results drawn according to expertise, the result finally using fuzzy synthesis Weight algorithm to draw goes to optimize the result that PSO BP algorithm draws, thus draw final classification results.Machine learning algorithm and traditional mathematical modelling algorithm are combined, the shortcoming having broken away from two kinds of algorithms, there is the unequal problem of classification when classification in machine learning algorithm, traditional mathematics modeling algorithm too relies on the experience level of expert, and new man-computer cooperation model can be good at solving the problem that both the above algorithm exists.

Description

A kind of man-computer cooperation credit evaluation new model towards medium-sized and small enterprises

Technical field

The present invention relates to medium-sized and small enterprises credit evaluation technical field, a kind of man-computer cooperation towards medium-sized and small enterprises is believed Borrow assessment new model.

Background technology

At present, owing to the economic development of country is played very important effect by medium-sized and small enterprises, the development of medium-sized and small enterprises is more More to be paid attention to by society and country, but, in the face of large-scale listed company of state-owned enterprise, medium-sized and small enterprises difficulty obtains existence from business bank The loan of fund, also rests on the policy-related loan stage in China for the loan of medium-sized and small enterprises at present, the most real Existing commercialization operates on a large scale.Until be interconnected net finance upsurge surround today, whether from national policy aspect still from Bring the biggest opportunity can in social environment the development of medium-sized and small enterprises.

It practice, the research of medium-sized and small enterprises credit evaluation is paid close attention to by Chinese scholars always.Some scholars use quantitatively Artificial intelligence approaches based on data medium-sized and small enterprises are carried out credit standing assessment, the most useful excellent based on mathematical modeling qualitatively Change method goes to be estimated medium-sized and small enterprises credit standing.Document 1 " Liu C, Xia X.The Credit Rating of Small and Medium Enterprises Based on Neural Network[C]//Information Engineering and Electronic Commerce(IEEC),2010 2nd International Symposium On.IEEE, 2010. " a kind of particle cluster algorithm optimization neural network evaluation model is proposed, and use matlab to emulate, Make this model more more efficient accurately than traditional BP neural network model in medium-sized and small enterprises credit evaluation.Document 2 " Wang H L.Credit Assessment for Listed Companies Based On GA-BP Model[C] .2010International Conference on E-Health Networking.2010. " connection revised by introducing The genetic algorithm of weighted value sets up BP neural network model, and chooses 2004 to 2007 80 listed companies as sample, But also constructing the credit evaluation financial ind ex system of a listed company, it is higher accurate that experiment shows that this model has Property.Document 3 " S.Zhu, X.Song, Y.Cao.Based on Multi-objective Fuzzy Decision Model to Study Tech Micro Enterprise Financing Credit Risk Evaluation System[C]// International Conference on Intelligent Transportation,Big Data and Smart City.IEEE, 2015. " it is based on Fuzzy Cluster Model, Medium-Small Sized Firms of High Technology is financed credit risk with grade scoring method Being assessed, according to the index weights of the weighted value obtained by calculating, carry out the analysis of result, it is medium and small according to new and high technology The feature of enterprise, determines level index, and refines two-level index, finally gives the result of fuzzy evaluation.Document 4 " Hsien- Chang Kuo,Lie-Huey Wang,Her-Jiun Sheu,et al.Credit Evaluation for Small and Medium-sized Enterprises by the Examination of Firm-specific Financial Ratios and Non-financial Variables:Evidence from Taiwan[J].Review of Pacific Basin Financial Markets&Policies, 2011,06 (01): 122-125. " in, the medium-sized and small enterprises credit to Taiwan is commented Introduce Non-financial when estimating, change the most single credit evaluation using financial data to carry out medium-sized and small enterprises.Document 5 “Zhang Y,Ji W G,Wang Y N.Modeling the credit risk for China's small and medium-sized enterprises[C].Emergency Management and Management Sciences (ICEMMS), 2011 2nd IEEE International Conference on.IEEE, 2011. " point out that new Basel assists View emphasizes the necessity of loan of small and medium enterprise portfolio credit risk model, finally propose eDplains model solve bank and in Credit risk problem between small enterprise.Document 6 " Hongli, Junchen.The credit rating of small and medium-sized enterprises based on the grey hierarchy evaluation model[C]// Information Science and Engineering(ICISE),20102nd International Conference On.2010. " think in that medium-sized and small enterprises credit evaluation can only obtain and incomplete has the highest grey information, this be due to The process of assessment is to be affected by factors such as the knowledge of expert, experience and preferences, and grey AHP assessment model can be for these Problem well processes, to this end, it is proposed that a kind of man-computer cooperation credit evaluation new model towards medium-sized and small enterprises.

Summary of the invention

It is an object of the invention to provide a kind of man-computer cooperation credit evaluation new model towards medium-sized and small enterprises, on solving State the problem proposed in background technology.

For achieving the above object, the present invention provides following technical scheme: a kind of man-computer cooperation credit towards medium-sized and small enterprises Assessment new model, comprises the following steps:

S1: initialize the inertial factor of PSO algorithm, Studying factors c₁, Studying factors c₂, population scale, greatest iteration time Number, the position of particle and the speed of particle；

S2: determine the topological structure of neutral net, the weights of BP neutral net and threshold value according to the dimension in data set Population maps, and data input pretreatment have an input layer, a hidden layer and the BP Neural Network of an output layer Network, can approach any one nonlinear function, therefore, the neutral net of one 3-tier architecture of structure with arbitrary accuracy；

S3: use the BP nerve in step S2 once to train, calculate n₁The training error E of individual training sample₁And n₂Individual Testing error E of test samples₂, in PSO-BP algorithm, the weights of BP network and threshold value more new formula is:

W_kj(t+1)=W_kj(t)+αδ_kH_j (1)

In above formula, H_jFor the output of hidden layer, W_kjT () is the hidden layer node j connection weights to output layer node k, α is for learning Practise parameter, generally 0.1～0.9, δ_kFor the error signal of output layer node k, error calculation formula is:

δ_k=(T_k-O_k)O_k(1-O_k) (2)

T_kFor the desired output of output node k, O_kFor real output value, O_kComputing formula is as follows:

$O_k=\underset{j=1}{\overset{s}{\Σ}}{W}_{kj}{H}_j+{\mathrm{\θ}}_k---\left(3\right)$

Wherein, s is hidden layer node number, and hidden layer output computing formula is:

$H_j=f(\underset{i=1}{\overset{M}{\Σ}}{W}_{ji}{I}_i+{\mathrm{\θ}}_j)---\left(4\right)$

W_ji(t+1)=W_ji(t)+ασ_jI_i (5)

In above formula, W_jiFor the connection weights of input layer i to hidden layer node j, I_iFor input layer i input Value, σ_jFor the error signal of hidden layer node j, error calculation formula is:

${\mathrm{\σ}}_j=\underset{k}{\Σ}{\mathrm{\δ}}_k{W}_{kj}{H}_j(1-H_j)---\left(6\right)$

θ_k(t+1)=θ_k(t)+βδ_k (7)

In above formula, θ_kFor the threshold value of output node k, β is learning parameter, generally 0.1～0.9,

θ_j(t+1)=θ_j(t)+βσ_j (8)

In above formula, θ_jFor the threshold value of hidden layer node j, shown

$E_1=\frac{1}{n_1}\underset{p_1=1}{\overset{n_1}{\Σ}}{(O_{p_1}-T_{p_1})}^2---\left(9\right)$

$E_2=\frac{1}{n_2}\underset{p_2=1}{\overset{n_2}{\Σ}}{(O_{p_2}-T_{p_2})}^2---\left(10\right)$

In above formula, n₁And n₂It is respectively training sample number and test samples number；WithIt is respectively training sample p₁ Network output and desired output,WithIt is respectively test samples p₂Network output and desired output；

S4: the weights of network are regarded as the speed of particle in PSO algorithm, and the change of twice weights is considered as the speed of particle in succession The change of degree, finds out minimum testing error E₂Time network weight, namely optimal adaptation value, in PSO algorithm to self speed and The formula of location updating is:

$v_{id}^{k+1}={\mathrm{wv}}_{id}^k+c_1{r}_{1d}^k\×({\mathrm{pBest}}_{id}^k-p_{id}^k)+c_2{r}_{2d}^k({\mathrm{gBest}}_d^k-p_{id}^k)---\left(11\right)$

$p_{id}^k=p_{id}^k+v_{id}^k---\left(12\right)$

Wherein, w is Inertia Weight, c₁、c₂For representing the constant of Studying factors, r_1d、r_2dFor the random number in [0,1], k is Iterations；

S5: often trained one time, considered the Joint effect of BP and PSO, will add by BP network weight adjustment formula The knots modification of weights；

S6: repeat above optimizing operation, the speed of more new particle and position.Stop condition is when reaching maximum appointment During evolution number of times, now can obtain global optimum's particle position and will be mapped to that weight and the Threshold-training of initial neutral net Neutral net；

S7: the optimum results drawn in step S6 is assigned to BP, is set the training of number of times, obtains a subseries knot Really；

S8: obtain comprehensive weight W according to comentropy and analytic hierarchy process (AHP)_jIn information Entropy Method, it is assumed that have n medium-sized and small enterprises Loan customer, interpretational criteria is C={C_i| i=1,2 ..., n}, then entropy weight can be defined as:

$\begin{array}{ccc}S(x_{ij}/\sqrt{\underset{i=1}{\overset{m}{\Σ}}{x}_{ij}^2})=-\underset{i=1}{\overset{m}{\Σ}}(x_{ij}/\sqrt{\underset{i=1}{\overset{m}{\Σ}}{x}_{ij}^2})\ln (x_{ij}/\sqrt{x_{ij}^2})& i=1,2,\mathrm{...}m;& j=1,\mathrm{2...},n.\end{array}---\left(13\right)$

$S_J=S(x_{ij}/\sqrt{\underset{i=1}{\overset{n}{\Σ}}{x}_{ij}^2})/\ln n---\left(14\right)$

G_J=1-S_J (15)

The comentropy weight summing up each criterion derived above is:

Analytic hierarchy structure is divided into destination layer A, rule layer B_k, solution layer C_n, destination layer only one of which, rule layer has k item Mesh, solution layer has n item, according to we can obtain the characteristic vector of B layer aboveFirst we first calculate each row element Product M_i, seek its n th RootThere is a following formula:

${\stackrel{\‾}{W}}_i=\sqrt[n]{M_i}---\left(17\right)$

WillDo normalization process to obtain:

$W_i=\frac{W_i}{\underset{j=1}{\overset{n}{\Σ}}{W}_j}---\left(18\right)$

Then W=[W₁ W₂...W_n]^rIt is exactly required characteristic vector, calculates the maximal eigenvector λ of matrix_maxAnd carry out one Cause is checked,

${\mathrm{\λ}}_{max}=\underset{i=1}{\overset{n}{\Σ}}\frac{{\left(AW\right)}_i}{{\mathrm{nW}}_i}---\left(19\right)$

$CI=\frac{{\mathrm{\λ}}_{max}-n}{n-1}---\left(20\right)$

$CR=\frac{CI}{RI}---\left(21\right)$

In like manner obtaining C layer weight vectors matrix, obtaining AHP weight is:

W_AHP=R₁W (22)

Comprehensive weight is can get by above (16) formula and (22) formula:

S9: use the fuzzy processing improved, try to achieve the credit value E of each medium-sized and small enterprises, in medium-sized and small enterprises credit evaluation There is certain fuzzy ambiguity, after the comprehensive weight of above two-level index calculates, tried to achieve by the method for Fuzzy Processing The credit value of each medium-sized and small enterprises.After fuzzy mathematics thought introduces, after using quantitative Treatment, needs assessment medium-sized and small enterprises provide Data constitute matrix Rj, replace the expert opinion matrix R in tradition expert judging method, show that computing formula is as follows:

E_j=R_j×W_j (24)

S10: use following Optimality Criteria to carry out the optimization of two times result:

1) if E_A>E_B, and in PSO-BP, A is bad enterprise, B enterprise preferably, then A has been classified as enterprise the most again；

2) if E_A<E_B, and A enterprise preferably in PSO-BP, B is bad enterprise, then A is classified as bad enterprise the most again；

S11: draw final optimum results more than comprehensive, provide medium-sized and small enterprises credit standing assessment result.

Compared with prior art, the invention has the beneficial effects as follows: towards the new mould of man-computer cooperation credit evaluation of medium-sized and small enterprises Type, combines machine learning algorithm and traditional mathematical modelling algorithm, the shortcoming having broken away from two kinds of algorithms, and machine learning is calculated There is the unequal problem of classification when classification in method, and traditional mathematics modeling algorithm too relies on the experience level of expert, and new people Machine combination model can be good at solving the problem that both the above algorithm exists.

Accompanying drawing explanation

Fig. 1 is the flow chart of the man-computer cooperation credit evaluation new model towards medium-sized and small enterprises of the present invention；

Fig. 2 is the PSO-BP algorithm fitness curve chart of the present invention.

Detailed description of the invention

Below in conjunction with the accompanying drawing in the embodiment of the present invention, the technical scheme in the embodiment of the present invention is carried out clear, complete Describe, it is clear that described embodiment is only a part of embodiment of the present invention rather than whole embodiments wholely.Based on Embodiment in the present invention, it is every other that those of ordinary skill in the art are obtained under not making creative work premise Embodiment, broadly falls into the scope of protection of the invention.

Referring to Fig. 1-2, the present invention provides a kind of technical scheme: a kind of man-computer cooperation credit evaluation towards medium-sized and small enterprises New model, comprises the following steps:

S1: initialize the inertial factor of PSO algorithm, Studying factors c₁, Studying factors c₂, population scale, greatest iteration time Number, the position of particle and the speed of particle；

S2: determine the topological structure of neutral net, the weights of BP neutral net and threshold value according to the dimension in data set Population maps, and data input pretreatment have an input layer, a hidden layer and the BP Neural Network of an output layer Network, can approach any one nonlinear function, therefore, the neutral net of one 3-tier architecture of structure with arbitrary accuracy；

S3: use the BP nerve in step S2 once to train, calculate n₁The training error E of individual training sample₁And n₂Individual Testing error E of test samples₂, in PSO-BP algorithm, the weights of BP network and threshold value more new formula is:

W_kj(t+1)=W_kj(t)+αδ_kH_j (1)

In above formula, H_jFor the output of hidden layer, W_kjT () is the hidden layer node j connection weights to output layer node k, α is for learning Practise parameter, generally 0.1～0.9, δ_kFor the error signal of output layer node k, error calculation formula is:

δ_k=(T_k-O_k)O_k(1-O_k) (2)

T_kFor the desired output of output node k, O_kFor real output value, O_kComputing formula is as follows:

$O_k=\underset{j=1}{\overset{s}{\&Sigma;}}{W}_{kj}{H}_j+{\mathrm{\&theta;}}_k---\left(3\right)$

Wherein, s is hidden layer node number, and hidden layer output computing formula is:

$H_j=f(\underset{i=1}{\overset{M}{\&Sigma;}}{W}_{ji}{I}_i+{\mathrm{\&theta;}}_j)---\left(4\right)$

W_ji(t+1)=W_ji(t)+ασ_jI_i (5)

In above formula, W_jiFor the connection weights of input layer i to hidden layer node j, I_iFor input layer i input Value, σ_jFor the error signal of hidden layer node j, error calculation formula is:

${\mathrm{\&sigma;}}_j=\underset{k}{\&Sigma;}{\mathrm{\&delta;}}_k{w}_{kj}{H}_j(1-H_j)---\left(6\right)$

θ_k(t+1)=θ_k(t)+βδ_k (7)

In above formula, θ_kFor the threshold value of output node k, β is learning parameter, generally 0.1～0.9,

θ_j(t+1)=θ_j(t)+βσ_j (8)

In above formula, θ_jFor the threshold value of hidden layer node j, shown

$E_1=\frac{1}{n_1}\underset{p_1=1}{\overset{n_1}{\&Sigma;}}{(O_{p_1}-T_{p_1})}^2---\left(9\right)$

$E_2=\frac{1}{n_2}\underset{p_2=1}{\overset{n_2}{\&Sigma;}}{(O_{p_2}-T_{p_2})}^2---\left(10\right)$

In above formula, n₁And n₂It is respectively training sample number and test samples number；WithIt is respectively training sample p₁ Network output and desired output,WithIt is respectively test samples p₂Network output and desired output；

S4: the weights of network are regarded as the speed of particle in PSO algorithm, and the change of twice weights is considered as the speed of particle in succession The change of degree, finds out minimum testing error E₂Time network weight, namely optimal adaptation value, in PSO algorithm to self speed and The formula of location updating is:

$v_{id}^{k+1}={\mathrm{wv}}_{id}^k+c_1{r}_{1d}^k\&times;({\mathrm{pBest}}_{id}^k-p_{id}^k)+c_2{r}_{2d}^k({\mathrm{gBest}}_d^k-p_{id}^k)---\left(11\right)$

$p_{id}^k=p_{id}^k+v_{id}^k---\left(12\right)$

Wherein, w is Inertia Weight, c₁、c₂For representing the constant of Studying factors, r_1d、r_2dFor the random number in [0,1], k is Iterations；

S5: often trained one time, considered the Joint effect of BP and PSO, will add by BP network weight adjustment formula The knots modification of weights；

S6: repeat above optimizing operation, the speed of more new particle and position.Stop condition is when reaching maximum appointment During evolution number of times, now can obtain global optimum's particle position and will be mapped to that weight and the Threshold-training of initial neutral net Neutral net；

S7: the optimum results drawn in step S6 is assigned to BP, is set the training of number of times, obtains a subseries knot Really；

S8: obtain comprehensive weight W according to comentropy and analytic hierarchy process (AHP)_jIn information Entropy Method, it is assumed that have n medium-sized and small enterprises Loan customer, interpretational criteria is C={C_i| i=1,2 ..., n}, then entropy weight can be defined as:

$\begin{array}{ccc}S(x_{ij}/\sqrt{\underset{i=1}{\overset{m}{\&Sigma;}}{x}_{ij}^2})=-\underset{i=1}{\overset{m}{\&Sigma;}}(x_{ij}/\sqrt{\underset{i=1}{\overset{m}{\&Sigma;}}{x}_{ij}^2})\ln (x_{ij}/\sqrt{x_{ij}^2})& i=1,2,\mathrm{...}m;& j=1,\mathrm{2...},n.\end{array}---\left(13\right)$

$S_J=S(x_{ij}/\sqrt{\underset{i=1}{\overset{n}{\&Sigma;}}{x}_{ij}^2})/\ln n---\left(14\right)$

G_J=1-S_J (15)

The comentropy weight summing up each criterion derived above is:

Analytic hierarchy structure is divided into destination layer A, rule layer B_k, solution layer C_n, destination layer only one of which, rule layer has k item Mesh, solution layer has n item, according to we can obtain the characteristic vector of B layer aboveFirst we first calculate each row element Product M_i, seek its n th RootThere is a following formula:

${\stackrel{\&OverBar;}{W}}_i=\sqrt[n]{M_i}---\left(17\right)$

WillDo normalization process to obtain:

$W_i=\frac{W_i}{\underset{j=1}{\overset{n}{\&Sigma;}}{W}_j}---\left(18\right)$

Then W=[W₁ W₂...W_n]^rIt is exactly required characteristic vector, calculates the maximal eigenvector λ of matrix_maxAnd carry out one Cause is checked,

${\mathrm{\&lambda;}}_{max}=\underset{i=1}{\overset{n}{\&Sigma;}}\frac{{\left(AW\right)}_i}{{\mathrm{nW}}_i}---\left(19\right)$

$CI=\frac{{\mathrm{\&lambda;}}_{max}-n}{n-1}---\left(20\right)$

$CR=\frac{CI}{RI}---\left(21\right)$

In like manner obtaining C layer weight vectors matrix, obtaining AHP weight is:

W_AHP=R₁W (22)

Comprehensive weight is can get by above (16) formula and (22) formula:

S9: use the fuzzy processing improved, try to achieve the credit value E of each medium-sized and small enterprises, in medium-sized and small enterprises credit evaluation There is certain fuzzy ambiguity, after the comprehensive weight of above two-level index calculates, tried to achieve by the method for Fuzzy Processing The credit value of each medium-sized and small enterprises.After fuzzy mathematics thought introduces, after using quantitative Treatment, needs assessment medium-sized and small enterprises provide Data constitute matrix Rj, replace the expert opinion matrix R in tradition expert judging method, show that computing formula is as follows:

E_j=R_j×W_j (24)

S10: use following Optimality Criteria to carry out the optimization of two times result:

1) if E_A>E_B, and in PSO-BP, A is bad enterprise, B enterprise preferably, then A has been classified as enterprise the most again；

2) if E_A<E_B, and A enterprise preferably in PSO-BP, B is bad enterprise, then A is classified as bad enterprise the most again；

S11: draw final optimum results more than comprehensive, provide medium-sized and small enterprises credit standing assessment result.

In view of the consideration to algorithm robustness, use three times of standard deviation methods of inspection to abnormal number in the legacy data obtained According to effectively rejecting, obtain initial data.Raw data set one has 32 attributes, by refering to document and related data with And four fiduciary loan index systems of big nationalized bank, use Best first search feature searching algorithm to evaluate letter with CFS Number finds out optimal feature subset, thus by 32 attributes according to Industry situation, financial risk, non-financial risk and situation of honouring an agreement Simplify 14 attributes.Consider national policy aspect and the parameter of industry development aspect first.The present invention at PSO-BP and In genetic-neural algorithm, two parts of data sets being carried out effective contrast verification, a copy of it does not contains country's Industry Policy and industry is sent out Exhibition trend attribute, the result finally drawn display uses the classification accuracy ratio that the data set experiment containing above two attributes is last The classification accuracy of the data set not containing above two attributes improves nearly 1%.

Although an embodiment of the present invention has been shown and described, for the ordinary skill in the art, permissible Understand and these embodiments can be carried out multiple change without departing from the principles and spirit of the present invention, revise, replace And modification, the scope of the present invention be defined by the appended.

Claims (1)

1. the man-computer cooperation credit evaluation new model towards medium-sized and small enterprises, it is characterised in that: comprise the following steps:

S1: initialize the inertial factor of PSO algorithm, Studying factors c₁, Studying factors c₂, population scale, maximum iteration time, grain The position of son and the speed of particle；

S2: determine the topological structure of neutral net, the weights of BP neutral net and the particle of threshold value according to the dimension in data set Group maps, and data input pretreatment have an input layer, a hidden layer and the feedforward neural network of an output layer, can To approach any one nonlinear function, therefore, the neutral net of one 3-tier architecture of structure with arbitrary accuracy；

S3: use the BP nerve in step S2 once to train, calculate n₁The training error E of individual training sample₁And n₂Individual inspection Testing error E of sample₂, in PSO-BP algorithm, the weights of BP network and threshold value more new formula is:

W_kj(t+1)=W_kj(t)+αδ_kH_j (1)

In above formula, H_jFor the output of hidden layer, W_kjT () is the hidden layer node j connection weights to output layer node k, α is study ginseng Number, generally 0.1～0.9, δ_kFor the error signal of output layer node k, error calculation formula is:

δ_k=(T_k-O_k)O_k(1-O_k) (2)

T_kFor the desired output of output node k, O_kFor real output value, O_kComputing formula is as follows:

Wherein, s is hidden layer node number, and hidden layer output computing formula is:

W_ji(t+1)=W_ji(t)+ασ_jI_i (5)

In above formula, W_jiFor the connection weights of input layer i to hidden layer node j, I_iFor the value of input layer i input, σ_jFor The error signal of hidden layer node j, error calculation formula is:

θ_k(t+1)=θ_k(t)+βδ_k (7)

In above formula, θ_kFor the threshold value of output node k, β is learning parameter, generally 0.1～0.9,

θ_j(t+1)=θ_j(t)+βσ_j (8)

In above formula, θ_jFor the threshold value of hidden layer node j, shown

In above formula, n₁And n₂It is respectively training sample number and test samples number；WithIt is respectively training sample p₁Network Output and desired output,WithIt is respectively test samples p₂Network output and desired output；

S4: the weights of network are regarded as the speed of particle in PSO algorithm, and the change of twice weights is considered as the speed of particle in succession Change, find out minimum testing error E₂Time network weight, namely optimal adaptation value, to self speed and position in PSO algorithm The formula updated is:

Wherein, w is Inertia Weight, c₁、c₂For representing the constant of Studying factors, r_1d、r_2dFor the random number in [0,1], k is iteration Number of times；

S5: often trained one time, considered the Joint effect of BP and PSO, will adjust formula plus weights by BP network weight Knots modification；

S6: repeat above optimizing operation, the speed of more new particle and position.Stop condition is when reaching maximum appointment evolution During number of times, now can obtain global optimum's particle position and will be mapped to that the weight of initial neutral net and Threshold-training are neural Network；

S7: the optimum results drawn in step S6 is assigned to BP, is set the training of number of times, obtains a classification results；

S8: obtain comprehensive weight W according to comentropy and analytic hierarchy process (AHP)_jIn information Entropy Method, it is assumed that have n loan of small and medium enterprise Client, interpretational criteria is C={C_i| i=1,2 ..., n}, then entropy weight can be defined as:

G_J=1-S_J (15)

The comentropy weight summing up each criterion derived above is:

Analytic hierarchy structure is divided into destination layer A, rule layer B_k, solution layer C_n, destination layer only one of which, rule layer has k project, Solution layer has n item, according to we can obtain the characteristic vector of B layer aboveFirst we first calculate each row element Product M_i, seek its n th RootThere is a following formula:

WillDo normalization process to obtain:

Then W=[W₁ W₂ ... W_n]^rIt is exactly required characteristic vector, calculates the maximal eigenvector λ of matrix_maxAnd carry out consistent Property inspection,

In like manner obtaining C layer weight vectors matrix, obtaining AHP weight is:

W_AHP=R₁W (22)

Comprehensive weight is can get by above (16) formula and (22) formula:

S9: use the fuzzy processing improved, try to achieve the credit value E of each medium-sized and small enterprises, exists in medium-sized and small enterprises credit evaluation Certain fuzzy ambiguity, after the comprehensive weight of above two-level index calculates, tries to achieve each by the method for Fuzzy Processing The credit value of medium-sized and small enterprises.After fuzzy mathematics thought introduces, after using quantitative Treatment, the data that needs assessment medium-sized and small enterprises provide Constitute matrix Rj, replace the expert opinion matrix R in tradition expert judging method, show that computing formula is as follows:

E_j=R_j×W_j (24)

S10: use following Optimality Criteria to carry out the optimization of two times result:

1) if E_A>E_B, and in PSO-BP, A is bad enterprise, B enterprise preferably, then A has been classified as enterprise the most again；

2) if E_A<E_B, and A enterprise preferably in PSO-BP, B is bad enterprise, then A is classified as bad enterprise the most again；

S11: draw final optimum results more than comprehensive, provide medium-sized and small enterprises credit standing assessment result.