Research On Credit Prediction Model Based On Deep Learning

Under the rapid development of Internet finance,various credit products have emerged.Such as Jingdong Finance,Ants Credit Pay,Ants By Chanting and other products.That has made it to be a very important research direction to recommend appropriate quotas for each user.Feature engineering always plays a key role in credit prediction or smart recommendation,but as the number and variety of features increases,the cost of creating,maintaining,and configuring artificially combined features is high.In the proposed deep learning-based model,there are their respective disadvantages in the feature extraction,the main research contents of this paper are as follows.(1)The cross operation of the original features in the Deep&Cross network limits the output to the special form of input,proposing a fusion cross network.Applying the vector dimension in the fusion cross network instead of the scalars in the Deep&Cross network.Redefining the calculation of each layer,unlike the simple crossing operation of the original feature in the Deep&Cross network,merging each layer The state of each layer in the fusion cross network is calculated from the state and original features of the previous layer.Experiments show that the results of fusion cross network on different data sets are better than deep cross network and so on.(2)In the neural network model,the problem of Degradation will occur as the number of layers deepens.This paper proposes two models based on the idea of residual network:Res-IN and Res-FC.Among them,the Res-IN model combines the residual unit with the sparse structural unit.Before training,the data is embedding into a matrix in the time dimension,and the convolution operation is used to capture the prediction relationship between the data;the Res-FC model is remove the convolution kernel in the residual unit and use the fully connected layer directly.This paper uses the Res-FC model when the data cannot be embedding into a matrix.Experiments on different data sets show that the two residual network-based models proposed in this paper have better experimental results.(3)In this paper,the fusion cross network is used to make explicit feature interaction,and the modified residual unit is used to make implicit feature interaction,which makes the feature generalization ability of the prediction model enhanced,and solves the drawbacks of explicit and implicit feature interaction.In order to increase the memory of the model,a linear unit is added to the model,so that the model also has excellent experimental results on a data set with a small amount of data.