Personal loan credit scoring method based on BA-WNN

Abstract

The invention discloses a personal loan credit scoring method based on BAWNN. The method comprises the following steps: S1, selecting parameters affecting credit scoring according to the information,obtained by an Internet financial platform, of a customer, and constructing a modeling sample set; s2, performing normalization processing on the modeling sample set, and then segmenting the trainingset and the test set; s3, initializing network parameters including the number of neurons of an input layer, a hidden layer and an output layer according to the training set, and constructing a wavelet neural network; s4, initializing a bat algorithm, optimizing wavelet neural network parameters by using a training sample and applying the bat algorithm, and training the optimized wavelet neural network to obtain a BAWNN prediction model; compared with other neural networks, the wavelet neural network not only has the local analysis characteristic of wavelet transform and the self-learning andself-adaptive capabilities of the neural network, but also can avoid blindness in structural design such as a BP neural network and the like, and has the advantages of simple network structure, high convergence speed, high precision and the like.

Description

technical field [0001] The invention belongs to the technical field of risk control in the Internet financial industry, and in particular relates to a BA-WNN-based personal loan credit scoring method. Background technique [0002] In recent years, the artificial neural network model has great advantages in the evaluation of Internet financial credit. It does not depend on specific assumptions in predicting, discovering and summarizing the structure of financial variables. At present, BP neural network, RBF neural network and Elman regression neural network are mostly used in credit evaluation applications, or some improvements based on these three networks. However, these three neural networks have various deficiencies when processing data for prediction: BP neural network is based on the gradient descent algorithm, which has defects such as local minima and low robustness; RBF neural network is a static feedforward network, which is There are deficiencies in dealing with d...

AI Technical Summary

Problems solved by technology

However, these three neural networks have various deficiencies when processing data for prediction: BP neural network is based on the gradient descent algorithm, which has defects such as local minima and low robustness; RBF neural network is a static feedforward network, which is There are deficiencies in dealing with dynamic time modeling problems; the Elman neural network is a local recurrent network, which has the disadvantages of many hidden units, slow convergence speed, and long training time.

[0003] Wavelet neural network (Wavelet Neural Network, WNN) is a kind of neural network that combines wavelet transform and neural network theory. Such as the Sigmoid function), organically combining the wavelet transform with the neural network, so that the wavelet neural network completely inherits the excellent time-frequency localization characteristics of the wavelet transform and the self-learning characteristics of the neural network, realizes the strong nonlinear approximation ability, and solves the problem of The traditional neural network prediction model has problems such as poor accuracy and low stability

[0004] However, because the wavelet neural network is the same as the BP neural network, the network parameters are gradually improved along the direction of local improvement through the gradient descent method, and it is easy to fall into the local extremum, and the convergence speed of the network is slow. optimization by weight

In recent years, genetic algorithm, particle swarm algorithm, and ant colony algorithm have been proposed to solve the problem of parameter optimization, but there are still some limitations in the nature of the target problem, parameter adjustment, and calculation time. The selection has a certain dependence. The particle swarm algorithm is difficult to select the maximum speed and the setting of the weighting factor, and the ant colony algorithm is difficult to select and set the parameters, and both have the disadvantages of being easy to fall into local optimum and slow convergence speed.

Images