CN106022517A - Risk prediction method and device based on nucleus limit learning machine - Google Patents

Abstract

The invention is suitable for the field of computers, provides a risk prediction method and device based on a nucleus limit learning machine and aims at solving problems in the prior art that the precision of risk prediction is not high because the penalty coefficient of the nucleus limit learning machine and the optimal value of nucleus width cannot be determined. The method comprises the steps: obtaining the operation data of a preset number of enterprises; carrying out the standardization of the operation data; optimizing the penalty coefficient of the nucleus limit learning machine and the nucleus width through employing a grey wolf algorithm, and obtaining the optimized penalty coefficient and nucleus width; constructing a prediction model of the nucleus limit learning machine based on the optimized penalty coefficient and nucleus width; and carrying out the risk prediction according to a prediction model. According to the technical scheme of the invention, the method integrates the grey wolf algorithm with the nucleus limit learning machine so as to determine the optimal values of the penalty coefficient and nucleus width, constructs the more accurate prediction model, achieves the effective prediction of risks, improves the prediction precision, and has a great application value for assisting a financial mechanism to carry out the scientific, reasonable and effective prediction of the operation risks of enterprises.

Description

A kind of method and apparatus of risk profile based on core extreme learning machine

Technical field

The present invention relates to field of computer technology, a kind of method particularly relating to risk profile based on core extreme learning machine And device.

Background technology

In order to reduce the loss that financial institution is caused by enterprise business risk especially clean risk of liquidation, have by setting up safety The Risk-warning mechanism of effect method that enterprise business risk is predicted, be financial institution keep investment repayment have efficacious prescriptions Formula.

Current existing enterprise business risk Forecasting Methodology can be divided mainly into two classes, i.e. based on statistical models method and Method based on artificial intelligence.Forecasting Methodology based on statistical models mainly have univariate analysis method, multivariate discriminant analysis method, Luo Jisite regression model and factor analysis.Compared to Forecasting Methodology based on statistical models, based on artificial intelligence pre- Survey method is widely used in Financial Risk Forecast field because of its superior performance.

Currently, Forecasting Methodology based on artificial intelligence mainly have based on artificial neural network, based on support vector machine, based on K neighbour, based on Bayesian model, based on extreme learning machine, based on the method such as mixed model and integrated study, these methods are all It is successfully applied to Financial Risk Forecast field.Wherein, method based on artificial neural network due to its can be preferably Catch the non-linear relation in data and applied receipts widely.But method based on artificial neural network is because generally using ladder Degree descent method carries out learning and there is the deficiency that is easily absorbed in local minimum, and network struction process simultaneously is also required to substantial amounts of ginseng Number is adjusted and has been difficult to set up the model of optimum.In order to overcome these shortcomings of neutral net, occur in that a kind of new god Through online learning methods i.e. extreme learning machine.Owing to extreme learning machine has good study generalization ability, learn based on the limit Model have begun to be applied to finance bankruptcy prediction, the prediction of economic life cycle and the prediction of rating business credit equivalent risk and comment In estimating.Owing to the input parameter value of extreme learning machine randomly generates, cause the performance of model to be not sufficiently stable, ask to solve this Topic, the concept of core extreme learning machine is suggested.Compared to extreme learning machine, core extreme learning machine need not be randomly provided input layer Weights with hidden layer, it is thus possible to obtain higher training speed, enterprise business risk based on core extreme learning machine is predicted Compare other Forecasting Methodologies more accurate.

But, now there are some researches show that the performance of core extreme learning machine is easily affected by two parameters, i.e. penalty coefficient and core width. Penalty coefficient for determine error of fitting minimize and input between balance weight minimize, core width defines from the input space Nonlinear mapping relation to high-dimensional feature space.The two key parameter needs predetermined, how to determine the punishment of optimum Coefficient and optimum core width, do not obtain preferably solution always, and at present, the method generally used is to utilize trellis search method true Determining their value, but network search method is easily absorbed in local optimum, the degree of accuracy thus resulting in risk profile is the highest.

Summary of the invention

The method and apparatus that it is an object of the invention to provide a kind of risk profile based on core extreme learning machine, it is intended to solve Certainly prior art cannot determine the penalty coefficient of core extreme learning machine and the optimum that core is wide, cause the degree of accuracy of risk profile The highest problem.

A first aspect of the present invention, it is provided that a kind of method of risk profile based on core extreme learning machine, including:

Obtaining the management data of the enterprise of predetermined quantity, described management data includes the feature of the attribute character of predetermined number Value；

Described management data is standardized；

Based on the described management data after standardization, utilize penalty coefficient and the core of grey wolf algorithm optimization core extreme learning machine Width, the penalty coefficient after being optimized and the core width after optimizing, the penalty coefficient after described optimization is used for determining that error of fitting is Balance weight between littleization and input data minimizes, and the core after described optimization is a width of empty from the input space to high dimensional feature Between nonlinear mapping relation；

The prediction of described core extreme learning machine is built based on the core width after the penalty coefficient after described optimization and described optimization Model；

Risk profile is carried out according to described forecast model.

A second aspect of the present invention, it is provided that the device of a kind of risk profile based on core extreme learning machine, including:

Data acquisition module, for obtaining the management data of the enterprise of predetermined quantity, described management data includes predetermined The eigenvalue of the attribute character of number；

Standardized module, for being standardized described management data；

Grey wolf optimizes module, for based on the described management data after standardization, utilizes the grey wolf algorithm optimization core limit Core width after the penalty coefficient of habit machine and core width, the penalty coefficient after being optimized and optimization, the penalty coefficient after described optimization For determining that the balance weight that error of fitting minimizes and input between data minimizes, the core after described optimization is a width of from defeated Enter the space nonlinear mapping relation to high-dimensional feature space；

Model construction module, for building described core based on the core width after the penalty coefficient after described optimization and described optimization The forecast model of extreme learning machine；

Prediction module, for carrying out risk profile according to described forecast model.

The present invention compared with prior art exists and provides the benefit that: by grey wolf algorithm incorporates core extreme learning machine Determine penalty coefficient and the core width of core extreme learning machine, due to grey wolf algorithm by natural imitation circle wolf pack social rank and Hunting behavior obtains the optimal solution of problem, therefore has more preferable search capability, it is possible to determine the punishment system of core extreme learning machine The optimum that number is wide with core, such that it is able to construct the forecast model of core extreme learning machine more accurately, it is achieved to enterprise's warp Effective prediction of battalion's risk, and substantially increase the degree of accuracy of prediction, in auxiliary, financial institution carries out section to enterprise business risk Learn in the most effective prediction and there is important using value.

Accompanying drawing explanation

Fig. 1 is the flow process of the method for a kind of based on core extreme learning machine the risk profile that the embodiment of the present invention one provides Figure；

Fig. 2 is the flow process of the method for a kind of based on core extreme learning machine the risk profile that the embodiment of the present invention two provides Figure；

Fig. 3 is that the structure of the device of a kind of based on core extreme learning machine the risk profile that the embodiment of the present invention three provides is shown It is intended to；

Fig. 4 is that the structure of the device of a kind of based on core extreme learning machine the risk profile that the embodiment of the present invention four provides is shown It is intended to.

Detailed description of the invention

In order to make the purpose of the present invention, technical scheme and advantage clearer, below in conjunction with drawings and Examples, right The present invention is further elaborated.Should be appreciated that specific embodiment described herein only in order to explain the present invention, and It is not used in the restriction present invention.

Below in conjunction with concrete accompanying drawing, the realization of the present invention is described in detail.

Embodiment one:

Fig. 1 is the flow process of the method for a kind of based on core extreme learning machine the risk profile that the embodiment of the present invention one provides Figure, specifically includes step S101 to S104, and details are as follows:

S101, obtaining the management data of enterprise of predetermined quantity, this management data includes the attribute character of predetermined number Eigenvalue.

Specifically, enterprise operation data refer to the attribute character of the enterprise's current economic situation from the definition of accounting angle, belong to Property feature represents a series of financial ratios, specifically can be, but not limited to include cash/current liability ratio (cash/current Liabilities), cash/total assets ratio (cash/total assets), current assets/flowing duty factor (current Assets/current liabilities), current assets/total assets ratio (current assets/total assets), battalion Fortune fund/total assets ratio (working capital/total assets), working capital/sales volume ratio (working Capital/sales), sales volume/stock's ratio (sales/inventory), sales volume/accounts receivable ratio (sales/ Receivables) etc..

In the enterprise of predetermined quantity, each enterprise provides the attribute character of predetermined number, these predetermined quantities The eigenvalue of the attribute character of the predetermined number of enterprise constitutes whole management datas.

S102, management data is standardized.

Specifically, in the management data of each enterprise, all it is standardized each attribute character processing so that Management data after standardization can support the realization of grey wolf algorithm accurately and effectively.

It is possible to further be standardized processing to management data according to formula (1):

$s_y{\[i\]}^{\′}=\frac{s_y\[i\]-s{\[i\]}_{\min }}{s{\[i\]}_{max}-s{\[i\]}_{min}}---\left(1\right)$

Wherein, s_y[i] ' is the eigenvalue after the ith attribute feature normalization of y-th enterprise, s_y[i] is y-th enterprise The eigenvalue of ith attribute feature, s [i]_maxFor the maximum of ith attribute feature of the enterprise of predetermined quantity, s [i]_min For the minima of ith attribute feature of the enterprise of predetermined quantity, y ∈ [1, Y], Y are the enterprise of predetermined quantity, i ∈ [1, I], I Attribute character for predetermined number.

S103, based on the management data after standardization, utilize grey wolf algorithm optimization core extreme learning machine penalty coefficient and Core width after core width, the penalty coefficient after being optimized and optimization, the penalty coefficient after optimization is used for determining that error of fitting is minimum Change and balance weight between input data minimizes, a width of non-from the input space to high-dimensional feature space of core after optimization Linear Mapping relation.

Specifically, the process of grey wolf algorithm be according to standardization after management data calculate the position of grey wolf, and determine row In the position of the grey wolf of front three, and according to the position of this front three grey wolf so that it is his grey wolf followed by these three fitst water Grey wolf go prey of surrounding and seize, i.e. allow the position of other grey wolves constantly surrounding and proximate to these three grey wolves, often update a position i.e. Completing an iteration process, and obtain a new location sets of wolf pack, iterating until reaching preset greatest iteration Till number of times.

The position of each grey wolf includes two parameters, i.e. penalty coefficient and core width, and penalty coefficient is used for determining that matching is by mistake Difference minimizes and inputs the balance weight between data and minimizes, a width of non-thread from the input space to high-dimensional feature space of core Property mapping relations.Therefore, in the final position of the wolf pack obtained by grey wolf algorithm is gathered, the grey wolf made number one i.e. head Two parameters of the position of wolf are the penalty coefficient after optimization and the core width after optimization.

S104, forecast model based on the core width structure core extreme learning machine after the penalty coefficient after optimizing and optimization.

Specifically, based on the core width after the penalty coefficient after the optimization that step S103 obtains and optimization, the core limit is carried out The structure of the forecast model of habit machine.

S105, carry out risk profile according to forecast model.

Specifically, according to the forecast model of the core extreme learning machine that step S104 builds, enterprise business risk can be entered Row prediction.

In the present embodiment, determine the penalty coefficient of core extreme learning machine by grey wolf algorithm is incorporated core extreme learning machine With core width, owing to grey wolf algorithm obtains the optimal solution of problem by the social rank of wolf pack in natural imitation circle and hunting behavior, Therefore there is more preferable search capability, it is possible to determine the penalty coefficient of core extreme learning machine and the optimum that core is wide, such that it is able to Construct the forecast model of core extreme learning machine more accurately, it is achieved the effective prediction to enterprise business risk, and significantly carry The high degree of accuracy of prediction, carries out having in scientific and reasonable effective prediction important in auxiliary financial institution to enterprise business risk Using value.

Embodiment two:

Fig. 2 is the flow process of the method for a kind of based on core extreme learning machine the risk profile that the embodiment of the present invention two provides Figure, specifically includes step S201 to S208, and details are as follows:

S201, obtaining the management data of enterprise of predetermined quantity, this management data includes the attribute character of predetermined number Eigenvalue.

Specifically, enterprise operation data refer to the attribute character of the enterprise's current economic situation from the definition of accounting angle, belong to Property feature represents a series of financial ratios, specifically can be, but not limited to include cash/current liability ratio (cash/current Liabilities), cash/total assets ratio (cash/total assets), current assets/flowing duty factor (current Assets/current liabilities), current assets/total assets ratio (current assets/total assets), battalion Fortune fund/total assets ratio (working capital/total assets), working capital/sales volume ratio (working Capital/sales), sales volume/stock's ratio (sales/inventory), sales volume/accounts receivable ratio (sales/ Receivables) etc..

In the enterprise of predetermined quantity, each enterprise provides the attribute character of predetermined number, these predetermined quantities The eigenvalue of the attribute character of the predetermined number of enterprise constitutes whole management data.

S202, management data is standardized.

In the management data of each enterprise, all it is standardized each attribute character processing so that standardization Management data after process can support the realization of grey wolf algorithm accurately and effectively.

Specifically, can be standardized processing to management data according to formula (2):

$s_y{\[i\]}^{\′}=\frac{s_y\[i\]-s{\[i\]}_{min}}{s{\[i\]}_{max}-s{\[i\]}_{min}}---\left(2\right)$

Wherein, s_y[i] ' is the eigenvalue after the ith attribute feature normalization of y-th enterprise, s_y[i] is y-th enterprise The eigenvalue of ith attribute feature, s [i]_maxFor the maximum of ith attribute feature of the enterprise of predetermined quantity, s [i]_min For the minima of ith attribute feature of the enterprise of predetermined quantity, y ∈ [1, Y], Y are the enterprise of predetermined quantity, i ∈ [1, I], I Attribute character for predetermined number.

S203, initializing grey wolf parameter, this grey wolf parameter includes maximum iteration time T, grey wolf number M, penalty coefficient C Hunting zone [C_min,C_max] and the hunting zone [γ of core width γ_min,γ_max]。

Specifically, to grey wolf algorithm needing the grey wolf parameter used initialize, can be, but not limited to include maximum Iterations T, grey wolf number M, the hunting zone [C of penalty coefficient C_min,C_max] and the hunting zone [γ of core width γ_min, γ_max]。

S204, the position X of each grey wolf is set_m, and set up grey wolf location matrix

Specifically, the position X of each grey wolf is set according to formula (3)_m:

X_m=(x_m1,x_m2) (3)

Wherein, x_m1For grey wolf the m value of penalty coefficient, x when current location_m2Wide for grey wolf m core when current location Value, m ∈ [1, M], x_m1∈[C_min,C_max], x_m2∈[γ_min,γ_max]。

Position X according to each grey wolf_m, build grey wolf location matrix according to formula (4)

$\stackrel{\→}{X}=\left[\begin{array}{ccc}{x}_{11}& & x_{12}\\ x_{21}& & x_{22}\\ & .& \\ & .& \\ & .& \\ x_{M1}& & x_{M2}\end{array}\right]---\left(4\right)$

S205, calculate the fitness f of each grey wolf_m。

Specifically, fitness f_mFor based on X_mAccuracy ACC of calculated core extreme learning machine, core extreme learning machine Accuracy ACC be after the management data after the standardization obtaining step S202 carries out K folding cross validation, acquisition average accurate Exactness, accuracy ACC of core extreme learning machine is calculated according to formula (5):

$ACC=\frac{{\mathrm{\Σ}}_1^K{\mathrm{acc}}_k}{K}---\left(5\right)$

Wherein, acc_kFor the accuracy of kth folding cross validation, K is the integer more than 0.

Cross validation (Cross-validation) refers in given modeling sample, takes out major part sample and carries out mould Type builds, and stays fraction sample to forecast the model set up, and seeks the prediction error of this fraction sample, this process Carry out, until all of sample has all been forecast once and only forecast once always.The purpose of cross validation is for terrible To reliable and stable model.K folding cross validation (K-fold cross-validation) is a kind of concrete shape of cross validation Formula, conventional K folding cross validation is 10 folding cross validations (10-fold cross validation).

In the present embodiment, K folding cross validation (K-fold cross-validation) refers to obtain step S202 Management data after standardization is divided into K subsample, and in turn using one of them subsample as test data, other K-1 are individual Subsample is as training data, it is ensured that each subsample is once as the chance of test data.Cross validation repeats K Secondary, each subsample is verified once, and the result of cross validation is acc each time_k, the result of average K cross validation obtains One single estimated value, i.e. accuracy ACC of core extreme learning machine.

S206, according to fitness f_mGrey wolf is ranked up, obtains the position with the grey wolf α of the highest fitnessTool There is the position of the grey wolf β of second highest fitnessPosition with the grey wolf δ of fitness three-hypers

Specifically, fitness f grey wolf obtained according to step S205_mValue be ranked up, obtain that there is the highest adaptation The position of the grey wolf α of degreeThere is the position of the grey wolf β of second highest fitnessPosition with the grey wolf δ of fitness three-hypersWherein there is the grey wolf α of the highest fitness as head wolf.

The distance of S207, respectively calculating grey wolf α, grey wolf β and grey wolf δ distance preyWith

Specifically, grey wolf α, grey wolf β and the distance of grey wolf δ distance prey are calculated according to formula (6) With

$\begin{array}{c}\stackrel{\→}{D_{\mathrm{\α}}}=\left|\stackrel{\→}{C_1}\·\stackrel{\→}{X_{\mathrm{\α}}}-\stackrel{\→}{X}\right|\\ \stackrel{\→}{D_{\mathrm{\β}}}=\left|\stackrel{\→}{C_2}\·\stackrel{\→}{X_{\mathrm{\β}}}-\stackrel{\→}{X}\right|\\ \stackrel{\→}{D_{\mathrm{\δ}}}=\left|\stackrel{\→}{C_3}\·\stackrel{\→}{X_{\mathrm{\δ}}}-\stackrel{\→}{X}\right|\end{array}---\left(6\right)$

Wherein,WithPass through formulaIt is calculated,For the random number between [0,1], i.e.Take difference Random number respectively obtainWith For current grey wolf location matrix,WithFor in current grey wolf position Put the position of the grey wolf coming front three in matrix.

S208, complete grey wolf location matrixRenewal.

Specifically, according to formula (7) to grey wolf location matrixIt is updated:

$\begin{array}{c}\stackrel{\→}{X}=\frac{\stackrel{\→}{X_1}+\stackrel{\→}{X_2}+\stackrel{\→}{X_3}}{3}\\ \stackrel{\→}{X_1}=\stackrel{\→}{X_{\mathrm{\α}}}-\stackrel{\→}{A_1}\·\left(\stackrel{\→}{D_{\mathrm{\α}}}\right)\\ \stackrel{\→}{X_2}=\stackrel{\→}{X_{\mathrm{\β}}}-\stackrel{\→}{A_2}\·\left(\stackrel{\→}{D_{\mathrm{\β}}}\right)\\ \stackrel{\→}{X_3}=\stackrel{\→}{X_{\mathrm{\δ}}}-\stackrel{\→}{A_3}\·\left(\stackrel{\→}{D_{\mathrm{\δ}}}\right)\end{array}---\left(7\right)$

Wherein,WithPass through formulaIt is calculated,T is current iteration time Number, t ∈ [1, T],Along with increase meeting linear decrease between 2 to 0 of iterations t,For the random number between [0,1], i.e.Take different randoms number to respectively obtainWith WithIt is calculated by formula (6).

The core concept of grey wolf algorithm be allow remove come other grey wolves of wolf pack front three followed by these three outstanding Grey wolf goes prey of surrounding and seize, and this is the algorithm core biology meaning of grey wolf algorithm.In specific algorithm, this core concept Simulated implementation is carried out by formula (7).Wherein,It is meant that to allow and wolf pack removes outside grey wolf α Other grey wolves followed by grey wolf α and remove hunting, and i.e. the position of other grey wolves is constantly close to grey wolf α,Implication Being to allow remove other grey wolves outside grey wolf β in wolf pack and followed by grey wolf β and remove hunting, i.e. the position of other grey wolves is constantly close to ash Wolf β,It is meant that to allow and wolf pack removes other grey wolves outside grey wolf δ followed by grey wolf δ and remove hunting, I.e. the position of other grey wolves is constantly close to grey wolf δ,Implication be then to round these three outstanding ashes Three grey wolf location matrixs of the new wolf pack that wolf producesWithIt is averaged, obtainedIt is this iteration The grey wolf location matrix of the wolf pack of rear renewal

S209, judge whether current iteration number of times t reaches described maximum iteration time T, if it is, perform step S210, otherwise according to the grey wolf location matrix after updatingReturn step S205 to continue executing with.

Specifically, if whether current iteration number of times t reaches described maximum iteration time T, then grey wolf has optimized, and performs Step S210, otherwise according to the grey wolf location matrix after updatingReturn step S205 and proceed iteration.

S210, the position X of output grey wolf α_α=(x_α1,x_α2), wherein x_α1For penalty coefficient C, x after optimizing_α2After optimizing Core width γ.

Specifically, when current iteration number of times t reaches described maximum iteration time T, the most up-to-date grey wolf location matrix In the parameter of the position of head wolf that makes number one be optimization after penalty coefficient C and core width γ after optimizing.I.e. grey wolf α's Position X_α=(x_α1,x_α2In), x_α1For penalty coefficient C, x after optimizing_α2For the core width γ after optimizing.

S211, forecast model based on the core width γ structure core extreme learning machine after the penalty coefficient C after optimizing and optimization.

Specifically, based on the core width γ after the penalty coefficient C after optimizing and optimization, the core limit is built according to formula (8) The forecast model of habit machine:

$\begin{array}{c}f\left(s\right)=\left[\begin{array}{c}K(s,s_1)\\ .\\ .\\ .\\ K(s,s_N)\end{array}\right]{(\frac{1}{C}+{\mathrm{\Ω}}_{ELM})}^{-1}T\\ K(s,s_n)=\exp (-\mathrm{\γ}||s-s_n|{|}^2)\end{array}---\left(8\right)$

Wherein, s is sample to be tested, s₁,…,s_NFor training sample, N is preset number of training, n ∈ [1, N], Ω_ELMFor the preset nuclear matrix meeting Mercer theorem structure.

Sample to be tested refers in the management data after the standardization that existing data set i.e. step S202 obtains, according to The rule preset takes a number of sample, and as the data of detection model performance, training sample refers to remove sample to be detected Other samples, as the training sample of training pattern.The number of samples of sample to be tested and training sample is by the value of N certainly Fixed.

Mercer theorem refers to that any positive semi-definite function can serve as kernel function.

S212, carry out risk profile according to forecast model.

Specifically, according to the forecast model of the core extreme learning machine that step S211 builds, enterprise business risk can be entered Row prediction.

In the present embodiment, determine the penalty coefficient of core extreme learning machine by grey wolf algorithm is incorporated core extreme learning machine With core width, specifically, penalty coefficient and core width are set as the location parameter of grey wolf and grey wolf location matrix is set, by calculating The value of fitness is also ranked up by the fitness of grey wolf, obtains the position that fitness comes the grey wolf of front three, and according to front Grey wolf location matrix is constantly updated in the grey wolf position of three so that it is the position of his grey wolf constantly close to these three grey wolves, is i.e. passed through In natural imitation circle, the social rank of wolf pack and hunting behavior obtain the optimal solution of problem, the i.e. penalty coefficient of core extreme learning machine The optimum wide with core, such that it is able to construct the forecast model of core extreme learning machine more accurately, it is achieved to enterprise operation Effective prediction of risk, and substantially increase the degree of accuracy of prediction, in auxiliary, financial institution carries out science to enterprise business risk The most effective prediction has important using value.

Embodiment three:

Fig. 3 is that the structure of the device of a kind of based on core extreme learning machine the risk profile that the embodiment of the present invention three provides is shown It is intended to, for convenience of description, illustrate only the part relevant to the embodiment of the present invention.The one of Fig. 3 example is based on the core limit The device of the risk profile of habit machine can be a kind of based on core extreme learning machine the risk profile that previous embodiment one provides The executive agent of method.The device concentrating risk profile based on core extreme learning machine of Fig. 3 example specifically includes that data acquisition Module 31, standardized module 32, grey wolf optimize module 33, model construction module 34 and prediction module 35.Each functional module is detailed It is described as follows:

Data acquisition module 31, for obtaining the management data of the enterprise of predetermined quantity, this management data includes predetermined The eigenvalue of the attribute character of number；

Standardized module 32, for being standardized management data；

Grey wolf optimizes module 33, for based on the management data after standardization, utilizes the grey wolf algorithm optimization core limit to learn Core width after the penalty coefficient of machine and core width, the penalty coefficient after being optimized and optimization, the penalty coefficient after optimization is for really Determining the balance weight that error of fitting minimizes and input between data to minimize, the core after optimization is a width of from the input space to height The nonlinear mapping relation of dimensional feature space；

Model construction module 34, for building core extreme learning machine based on the core width after the penalty coefficient after optimizing and optimization Forecast model；

Prediction module 35, for carrying out risk profile according to forecast model.

In the device of a kind of based on core extreme learning machine the risk profile that the present embodiment provides, each module realizes respective merit The process of energy, specifically refers to the description of aforementioned embodiment illustrated in fig. 1, and here is omitted.

Knowable to the device of a kind of based on core extreme learning machine the risk profile of above-mentioned Fig. 3 example, in the present embodiment, logical Cross penalty coefficient and core width grey wolf algorithm being incorporated core extreme learning machine to determine core extreme learning machine, owing to grey wolf algorithm leads to Cross social rank and the optimal solution of hunting behavior acquisition problem of wolf pack in natural imitation circle, therefore have and preferably search for energy Power, it is possible to determine the penalty coefficient of core extreme learning machine and the optimum that core is wide, such that it is able to construct core pole more accurately The forecast model of limit learning machine, it is achieved the effective prediction to enterprise business risk, and substantially increase the degree of accuracy of prediction, auxiliary Help financial institution that enterprise business risk carries out have important using value in scientific and reasonable effective prediction.

Embodiment four:

Fig. 4 is that the structure of the device of a kind of based on core extreme learning machine the risk profile that the embodiment of the present invention four provides is shown It is intended to, for convenience of description, illustrate only the part relevant to the embodiment of the present invention.The one of Fig. 4 example is based on the core limit The device of the risk profile of habit machine can be a kind of based on core extreme learning machine the risk profile that previous embodiment two provides The executive agent of method.The device concentrating risk profile based on core extreme learning machine of Fig. 4 example specifically includes that data acquisition Module 41, standardized module 42, grey wolf optimize module 43, model construction module 44 and prediction module 45.Each functional module is detailed It is described as follows:

Data acquisition module 41, for obtaining the management data of the enterprise of predetermined quantity, this management data includes predetermined The eigenvalue of the attribute character of number；

Standardized module 42, for being standardized management data；

Grey wolf optimizes module 43, for based on the management data after standardization, utilizes the grey wolf algorithm optimization core limit to learn Core width after the penalty coefficient of machine and core width, the penalty coefficient after being optimized and optimization, the penalty coefficient after optimization is for really Determining the balance weight that error of fitting minimizes and input between data to minimize, the core after optimization is a width of from the input space to height The nonlinear mapping relation of dimensional feature space；

Model construction module 44, for building core extreme learning machine based on the core width after the penalty coefficient after optimizing and optimization Forecast model；

Prediction module 45, for carrying out risk profile according to forecast model.

Further, standardized module 42, it is additionally operable to:

As follows management data is standardized:

$s_y{\[i\]}^{\′}=\frac{s_y\[i\]-s{\[i\]}_{min}}{s{\[i\]}_{max}-s{\[i\]}_{min}}$

Wherein, s_y[i] ' is the eigenvalue after the ith attribute feature normalization of y-th enterprise, s_y[i] is y-th enterprise The eigenvalue of ith attribute feature, s [i]_maxFor the maximum of ith attribute feature of the enterprise of described predetermined quantity, s [i]_minFor the minima of ith attribute feature of the enterprise of described predetermined quantity, y ∈ [1, Y], Y are the enterprise of described predetermined quantity Industry, i ∈ [1, I], I are the attribute character of described predetermined number.

Further, grey wolf optimization module 43 includes:

Initialization submodule 431, is used for initializing grey wolf parameter, and this grey wolf parameter includes maximum iteration time T, grey wolf Number M, the hunting zone [C of penalty coefficient C_min,C_max] and the hunting zone [γ of core width γ_min,γ_max]；

Position arranges submodule 432, for arranging the position X of each grey wolf as follows_m, and set up grey wolf position Matrix

X_m=(x_m1,x_m2)

$\stackrel{\→}{X}=\left[\begin{array}{ccc}{x}_{11}& & x_{12}\\ x_{21}& & x_{22}\\ & .& \\ & .& \\ & .& \\ x_{M1}& & x_{M2}\end{array}\right]$

Wherein, x_m1For grey wolf the m value of penalty coefficient, x when current location_m2Wide for grey wolf m core when current location Value, m ∈ [1, M], x_m1∈[C_min,C_max], x_m2∈[γ_min,γ_max]；

Optimized Iterative submodule 433, for calculating the fitness of grey wolf, and according to suitable according to the management data after standardization Grey wolf is ranked up by response, updates the grey wolf location matrix after grey wolf position is updatedComplete a grey wolf to optimize repeatedly Generation, and according to the grey wolf location matrix after updatingProceeding grey wolf Optimized Iterative, until reaching maximum iteration time T being Only, current grey wolf location matrix is exportedThe position of middle head wolf, penalty coefficient C after being optimized according to the position of head wolf and excellent Core width γ after change.

Further, Optimized Iterative submodule 433, it is additionally operable to:

Calculate the fitness f of each grey wolf_m, this fitness f_mFor based on X_mThe standard of calculated core extreme learning machine Exactness ACC, accuracy ACC is the bat that K based on the management data after standardization folding cross validation obtains, accuracy The computing formula of ACC is:

$ACC=\frac{{\mathrm{\Σ}}_1^K{\mathrm{acc}}_k}{K}$

Wherein, acc_kFor the accuracy of kth folding cross validation, K is the integer more than 0；

According to fitness f_mGrey wolf is ranked up, obtains the position with the grey wolf α of the highest fitnessHave secondary The position of the grey wolf β of high fitnessPosition with the grey wolf δ of fitness three-hypers

Grey wolf α, grey wolf β and the distance of grey wolf δ distance prey is calculated respectively according to equation belowWith

$\stackrel{\→}{D_{\mathrm{\α}}}=\left|\stackrel{\→}{C_1}\·\stackrel{\→}{X_{\mathrm{\α}}}-\stackrel{\→}{X}\right|$

$\stackrel{\→}{D_{\mathrm{\β}}}=\left|\stackrel{\→}{C_2}\·\stackrel{\→}{X_{\mathrm{\β}}}-\stackrel{\→}{X}\right|$

$\stackrel{\→}{D_{\mathrm{\δ}}}=\left|\stackrel{\→}{C_3}\·\stackrel{\→}{X_{\mathrm{\δ}}}-\stackrel{\→}{X}\right|$

Wherein,WithPass through formulaIt is calculated,For the random number between [0,1]；

Complete grey wolf location matrix according to equation belowRenewal:

$\stackrel{\→}{X}=\frac{\stackrel{\→}{X_1}+\stackrel{\→}{X_2}+\stackrel{\→}{X_3}}{3}$

$\stackrel{\→}{X_1}=\stackrel{\→}{X_{\mathrm{\α}}}-\stackrel{\→}{A_1}\·\left(\stackrel{\→}{D_{\mathrm{\α}}}\right)$

$\stackrel{\→}{X_2}=\stackrel{\→}{X_{\mathrm{\β}}}-\stackrel{\→}{A_2}\·\left(\stackrel{\→}{D_{\mathrm{\β}}}\right)$

$\stackrel{\→}{X_3}=\stackrel{\→}{X_{\mathrm{\δ}}}-\stackrel{\→}{A_3}\·\left(\stackrel{\→}{D_{\mathrm{\δ}}}\right)$

Wherein,WithPass through formulaIt is calculated,T is current iteration time Number, t ∈ [1, T],For the random number between [0,1]；

If current iteration number of times t not yet reaches maximum iteration time T, then according to the grey wolf location matrix after updatingContinue Continue and carry out grey wolf Optimized Iterative, until current iteration number of times t reaches maximum iteration time T；

The position X of output grey wolf α_α=(x_α1,x_α2), wherein x_α1For penalty coefficient C, x after optimizing_α2For the core after optimizing Wide γ.

Further, model construction module 44 includes:

Forecast model builds submodule 441, the core width γ after being used for based on the penalty coefficient C after optimizing and optimization, according to The forecast model of equation below structure core extreme learning machine:

$f\left(s\right)=\left[\begin{array}{c}K(s,s_1)\\ .\\ .\\ .\\ K(s,s_N)\end{array}\right]{(\frac{1}{C}+{\mathrm{\Ω}}_{ELM})}^{-1}T$

K(s,s_n)=exp (-γ | | s-s_n||²)

Wherein, s is sample to be tested, s₁,…,s_NFor training sample, N is preset number of training, n ∈ [1, N], Ω_ELMFor the preset nuclear matrix meeting Mercer theorem structure.

In the device of a kind of based on core extreme learning machine the risk profile that the present embodiment provides, each module realizes respective merit The process of energy, specifically refers to the description of aforementioned embodiment illustrated in fig. 2, and here is omitted.

Knowable to the device of a kind of based on core extreme learning machine the risk profile of above-mentioned Fig. 4 example, in the present embodiment, logical Cross penalty coefficient and core width grey wolf algorithm being incorporated core extreme learning machine to determine core extreme learning machine, specifically, will punish Coefficient and core width are set as the location parameter of grey wolf and arrange grey wolf location matrix, by calculating the fitness of grey wolf and to adaptation The value of degree is ranked up, and obtains the position that fitness comes the grey wolf of front three, and the grey wolf position according to front three is continuous the most more New grey wolf location matrix so that it is the position of his grey wolf is constantly close to these three grey wolves, i.e. by the society of wolf pack in natural imitation circle Meeting grade and hunting behavior obtain the penalty coefficient of the optimal solution of problem, i.e. core extreme learning machine and the optimum that core is wide, thus The forecast model of core extreme learning machine more accurately can be constructed, it is achieved the effective prediction to enterprise business risk, and greatly Improve greatly the degree of accuracy of prediction, assisting financial institution that enterprise business risk is carried out, scientific and reasonable effective prediction has Important using value.

It should be noted that each embodiment in this specification all uses the mode gone forward one by one to describe, each embodiment Stress is all the difference with other embodiments, and between each embodiment, same or like part sees mutually ?.For device class embodiment, due to itself and embodiment of the method basic simlarity, so describe is fairly simple, relevant Part sees the part of embodiment of the method and illustrates.

It should be noted that in said apparatus embodiment, included modules simply carries out drawing according to function logic Point, but it is not limited to above-mentioned division, as long as being capable of corresponding function；It addition, each functional module is concrete Title also only to facilitate mutually distinguish, is not limited to protection scope of the present invention.

It will appreciated by the skilled person that all or part of step realizing in the various embodiments described above method is can Completing instructing relevant hardware by program, corresponding program can be stored in a computer read/write memory medium In, described storage medium, such as ROM/RAM, disk or CD etc..

The foregoing is only presently preferred embodiments of the present invention, not in order to limit the present invention, all essences in the present invention Any amendment, equivalent and the improvement etc. made within god and principle, should be included within the scope of the present invention.

Claims (10)

1. the method for a risk profile based on core extreme learning machine, it is characterised in that including:

Obtaining the management data of the enterprise of predetermined quantity, described management data includes the eigenvalue of the attribute character of predetermined number；

Described management data is standardized；

Based on the described management data after standardization, utilize penalty coefficient and the core width of grey wolf algorithm optimization core extreme learning machine, Penalty coefficient after being optimized and the core width after optimization, the penalty coefficient after described optimization is used for determining that error of fitting minimizes And the balance weight between input data minimizes, core after described optimization is a width of from the input space to high-dimensional feature space Nonlinear mapping relation；

The forecast model of described core extreme learning machine is built based on the core width after the penalty coefficient after described optimization and described optimization；

Risk profile is carried out according to described forecast model.

The method of risk profile based on core extreme learning machine the most according to claim 1, it is characterised in that described to institute State management data to be standardized including:

As follows described management data is standardized:

$s_y{\[i\]}^{\′}=\frac{s_y\[i\]-s{\[i\]}_{min}}{s{\[i\]}_{max}-s{\[i\]}_{min}}$

Wherein, s_y[i] ' is the eigenvalue after the ith attribute feature normalization of y-th enterprise, s_y[i] is the of y-th enterprise The eigenvalue of i attribute character, s [i]_maxFor the maximum of ith attribute feature of the enterprise of described predetermined quantity, s [i]_min For the minima of ith attribute feature of the enterprise of described predetermined quantity, y ∈ [1, Y], Y are the enterprise of described predetermined quantity, i ∈ [1, I], I are the attribute character of described predetermined number.

The method of risk profile based on core extreme learning machine the most according to claim 1, it is characterised in that described based on Described management data after standardization, utilizes penalty coefficient and the core width of grey wolf algorithm optimization core extreme learning machine, is optimized After penalty coefficient and optimize after core width include:

Initializing grey wolf parameter, described grey wolf parameter includes the search model of maximum iteration time T, grey wolf number M, penalty coefficient C Enclose [C_min,C_max] and the hunting zone [γ of core width γ_min,γ_max]；

The position X of each grey wolf is set as follows_m, and set up grey wolf location matrix

X_m=(x_m1,x_m2)

$\stackrel{\→}{X}=\left[\begin{array}{ccc}{x}_{11}& & x_{12}\\ x_{21}& & x_{22}\\ & .& \\ & .& \\ & .& \\ x_{M1}& & x_{M2}\end{array}\right]$

Wherein, x_m1For grey wolf the m value of described penalty coefficient, x when current location_m2For grey wolf m described core width when current location Value, m ∈ [1, M], x_m1∈[C_min,C_max], x_m2∈[γ_min,γ_max]；

Calculate the fitness of grey wolf according to the described management data after standardization, and according to described fitness, grey wolf is arranged Sequence, updates the grey wolf location matrix after grey wolf position is updatedComplete a grey wolf Optimized Iterative, and according to described renewal After grey wolf location matrixProceed described grey wolf Optimized Iterative, until reaching described maximum iteration time T, output Current grey wolf location matrixThe position of middle head wolf, the penalty coefficient C after being optimized according to the position of described head wolf and optimization After core width γ.

The method of risk profile based on core extreme learning machine the most according to claim 3, it is characterised in that described basis Described management data after standardization calculates the fitness of grey wolf, and is ranked up grey wolf according to described fitness, updates ash Wolf position updated after grey wolf location matrixComplete a grey wolf Optimized Iterative, and according to the grey wolf after described renewal Location matrixProceeding described grey wolf Optimized Iterative, until reaching described maximum iteration time T, exporting current grey wolf Location matrixCore width after the position of middle head wolf, the penalty coefficient C after being optimized according to the position of described head wolf and optimization γ includes:

Calculate the fitness f of described each grey wolf_m, described fitness f_mFor based on X_mThe calculated described core limit learns Accuracy ACC of machine, described accuracy ACC is the flat of K based on the described management data after standardization folding cross validation acquisition All accuracy, the computing formula of described accuracy ACC is:

$ACC=\frac{{\mathrm{\Σ}}_1^K{\mathrm{acc}}_k}{K}$

Wherein, acc_kFor the accuracy of kth folding cross validation, K is the integer more than 0；

According to described fitness f_mGrey wolf is ranked up, obtains the position with the grey wolf α of the highest fitnessHave second highest The position of the grey wolf β of fitnessPosition with the grey wolf δ of fitness three-hypers

Described grey wolf α, described grey wolf β and the distance of described grey wolf δ distance prey is calculated respectively according to equation below With

$\stackrel{\→}{D_{\mathrm{\α}}}=|\stackrel{\→}{C_1}\·\stackrel{\→}{X_{\mathrm{\α}}}-\stackrel{\→}{X}|$

$\stackrel{\→}{D_{\mathrm{\β}}}=|\stackrel{\→}{C_2}\·\stackrel{\→}{X_{\mathrm{\β}}}-\stackrel{\→}{X}|$

$\stackrel{\→}{D_{\mathrm{\δ}}}=|\stackrel{\→}{C_3}\·\stackrel{\→}{X_{\mathrm{\δ}}}-\stackrel{\→}{X}|$

Wherein,WithPass through formulaIt is calculated,For the random number between [0,1]；

Complete grey wolf location matrix according to equation belowRenewal:

$\stackrel{\→}{X}=\frac{\stackrel{\→}{X_1}+\stackrel{\→}{X_2}+\stackrel{\→}{X_3}}{3}$

$\stackrel{\→}{X_1}=\stackrel{\→}{X_{\mathrm{\α}}}-\stackrel{\→}{A_1}\·\left(\stackrel{\→}{D_{\mathrm{\α}}}\right)$

$\stackrel{\→}{X_2}=\stackrel{\→}{X_{\mathrm{\β}}}-\stackrel{\→}{A_2}\·\left(\stackrel{\→}{D_{\mathrm{\β}}}\right)$

$\stackrel{\→}{X_3}=\stackrel{\→}{X_{\mathrm{\δ}}}-\stackrel{\→}{A_3}\·\left(\stackrel{\→}{D_{\mathrm{\δ}}}\right)$

Wherein,WithPass through formulaIt is calculated,T is current iteration number of times, t ∈ [1, T],For the random number between [0,1]；

If described current iteration number of times t not yet reaches described maximum iteration time T, then according to the grey wolf location matrix after updatingProceed grey wolf Optimized Iterative, until described current iteration number of times t reaches described maximum iteration time T；

The position X of output grey wolf α_α=(x_α1,x_α2), wherein x_α1For penalty coefficient C, x after optimizing_α2For the core width γ after optimizing.

The method of risk profile based on core extreme learning machine the most according to claim 1, it is characterised in that described based on Penalty coefficient after described optimization and the core width after described optimization build the forecast model of described core extreme learning machine and include:

Based on the penalty coefficient C after described optimization and the core width γ after described optimization, build the described core limit according to equation below The forecast model of learning machine:

$f\left(s\right)=\left[\begin{array}{c}K(s,s_1)\\ .\\ .\\ .\\ K(s,s_N)\end{array}\right]{(\frac{1}{C}+{\mathrm{\Ω}}_{ELM})}^{-1}T$

K(s,s_n)=exp (-γ | | s-s_n||²)

Wherein, s is sample to be tested, s₁,…,s_NFor training sample, N is preset number of training, n ∈ [1, N], Ω_ELMFor The preset nuclear matrix meeting Mercer theorem structure.

6. the device of a risk profile based on core extreme learning machine, it is characterised in that including:

Data acquisition module, for obtaining the management data of the enterprise of predetermined quantity, described management data includes predetermined number The eigenvalue of attribute character；

Standardized module, for being standardized described management data；

Grey wolf optimizes module, for based on the described management data after standardization, utilizes grey wolf algorithm optimization core extreme learning machine Penalty coefficient and core width, the penalty coefficient after being optimized and optimize after core width, the penalty coefficient after described optimization is used for Determining that the balance weight that error of fitting minimizes and input between data minimizes, the core after described optimization is a width of from input sky Between to the nonlinear mapping relation of high-dimensional feature space；

Model construction module, for building the described core limit based on the core width after the penalty coefficient after described optimization and described optimization The forecast model of learning machine；

Prediction module, for carrying out risk profile according to described forecast model.

The device of risk profile based on core extreme learning machine the most according to claim 6, it is characterised in that described standard Change module to include:

Standardization submodule, for as follows described management data being standardized:

$s_y{\[i\]}^{\′}=\frac{s_y\[i\]-s{\[i\]}_{min}}{s{\[i\]}_{max}-s{\[i\]}_{min}}$

Wherein, s_y[i] ' is the eigenvalue after the ith attribute feature normalization of y-th enterprise, s_y[i] is the of y-th enterprise The eigenvalue of i attribute character, s [i]_maxFor the maximum of ith attribute feature of the enterprise of described predetermined quantity, s [i]_min For the minima of ith attribute feature of the enterprise of described predetermined quantity, y ∈ [1, Y], Y are the enterprise of described predetermined quantity, i ∈ [1, I], I are the attribute character of described predetermined number.

The device of risk profile based on core extreme learning machine the most according to claim 6, it is characterised in that described grey wolf Optimize module to include:

Initialization submodule, is used for initializing grey wolf parameter, described grey wolf parameter include maximum iteration time T, grey wolf number M, Hunting zone [the C of penalty coefficient C_min,C_max] and the hunting zone [γ of core width γ_min,γ_max]；

Position arranges submodule, for arranging the position X of each grey wolf as follows_m, and set up grey wolf location matrix

X_m=(x_m1,x_m2)

$\stackrel{\→}{X}=\left[\begin{array}{ccc}{x}_{11}& & x_{12}\\ x_{21}& & x_{22}\\ & .& \\ & .& \\ & .& \\ x_{M1}& & x_{M2}\end{array}\right]$

Wherein, x_m1For grey wolf the m value of described penalty coefficient, x when current location_m2For grey wolf m described core width when current location Value, m ∈ [1, M], x_m1∈[C_min,C_max], x_m2∈[γ_min,γ_max]；

Optimized Iterative submodule, for calculating the fitness of grey wolf, and according to described according to the described management data after standardization Grey wolf is ranked up by fitness, updates the grey wolf location matrix after grey wolf position is updatedComplete a grey wolf to optimize Iteration, and according to the grey wolf location matrix after described renewalProceed described grey wolf Optimized Iterative, until described in reaching Till big iterations T, export current grey wolf location matrixThe position of middle head wolf, is optimized according to the position of described head wolf After penalty coefficient C and optimize after core width γ.

The device of risk profile based on core extreme learning machine the most according to claim 8, it is characterised in that described optimization Iteration submodule, is additionally operable to:

Calculate the fitness f of described each grey wolf_m, described fitness f_mFor based on X_mThe calculated described core limit learns Accuracy ACC of machine, described accuracy ACC is the flat of K based on the described management data after standardization folding cross validation acquisition All accuracy, the computing formula of described accuracy ACC is:

$ACC=\frac{{\mathrm{\Σ}}_1^K{\mathrm{acc}}_k}{K}$

Wherein, acc_kFor the accuracy of kth folding cross validation, K is the integer more than 0；

According to described fitness f_mGrey wolf is ranked up, obtains the position with the grey wolf α of the highest fitnessHave second highest The position of the grey wolf β of fitnessPosition with the grey wolf δ of fitness three-hypers

Described grey wolf α, described grey wolf β and the distance of described grey wolf δ distance prey is calculated respectively according to equation below With

$\stackrel{\→}{D_{\mathrm{\α}}}=|\stackrel{\→}{C_1}\·\stackrel{\→}{X_{\mathrm{\α}}}-\stackrel{\→}{X}|$

$\stackrel{\→}{D_{\mathrm{\β}}}=|\stackrel{\→}{C_2}\·\stackrel{\→}{X_{\mathrm{\β}}}-\stackrel{\→}{X}|$

$\stackrel{\→}{D_{\mathrm{\δ}}}=|\stackrel{\→}{C_3}\·\stackrel{\→}{X_{\mathrm{\δ}}}-\stackrel{\→}{X}|$

Wherein,WithPass through formulaIt is calculated,For the random number between [0,1]；

Complete grey wolf location matrix according to equation belowRenewal:

$\stackrel{\→}{X}=\frac{\stackrel{\→}{X_1}+\stackrel{\→}{X_2}+\stackrel{\→}{X_3}}{3}$

$\stackrel{\→}{X_1}=\stackrel{\→}{X_{\mathrm{\α}}}-\stackrel{\→}{A_1}\·\left(\stackrel{\→}{D_{\mathrm{\α}}}\right)$

$\stackrel{\→}{X_2}=\stackrel{\→}{X_{\mathrm{\β}}}-\stackrel{\→}{A_2}\·\left(\stackrel{\→}{D_{\mathrm{\β}}}\right)$

$\stackrel{\→}{X_3}=\stackrel{\→}{X_{\mathrm{\δ}}}-\stackrel{\→}{A_3}\·\left(\stackrel{\→}{D_{\mathrm{\δ}}}\right)$

Wherein,WithPass through formulaIt is calculated,T is current iteration number of times, t ∈ [1, T],For the random number between [0,1]；

If described current iteration number of times t not yet reaches described maximum iteration time T, then according to the grey wolf location matrix after updatingProceed grey wolf Optimized Iterative, until described current iteration number of times t reaches described maximum iteration time T；

The position X of output grey wolf α_α=(x_α1,x_α2), wherein x_α1For penalty coefficient C, x after optimizing_α2For the core width γ after optimizing.

The device of risk profile based on core extreme learning machine the most according to claim 6, it is characterised in that described mould Type builds module and includes:

Forecast model builds submodule, for based on the penalty coefficient C after described optimization and the core width γ after described optimization, according to The forecast model of the equation below described core extreme learning machine of structure:

$f\left(s\right)=\left[\begin{array}{c}K(s,s_1)\\ .\\ .\\ .\\ K(s,s_N)\end{array}\right]{(\frac{1}{C}+{\mathrm{\Ω}}_{ELM})}^{-1}T$

K(s,s_n)=exp (-γ | | s-s_n||²)

Wherein, s is sample to be tested, s₁,…,s_NFor training sample, N is preset number of training, n ∈ [1, N], Ω_ELMFor The preset nuclear matrix meeting Mercer theorem structure.