Research On Online Lending Customer Behavior Prediction Based On Deep Learning

In recent years,with the continuous development of Chinese economy and society,as well as the rapid rise of Internet finance,the online lending business of financial institutions has gradually emerged and the volume of lending transactions through the Internet has increased year by year.While internet lending has brought huge profits to financial institutions,it has also posed a huge challenge to their internet finance risk control capabilities.Accurate assessment of borrowing customers’ repayment behavior is a core element of Internet finance risk control.This research predicted two customer repayment behaviors(e.g.default repayment and prepayment)by applying improved Deep Residual Network.The details of this research are as follows:(1)The complex,high-dimensional,and noisy modern economic data,have led to the poor performance of model based on machine learning and deep learning methods,in predicting customer repayment behaviors.To address the issues mentioned above,based on Res Net model,integrating multi-level features and SENet,this research proposed the IRes Net(Improved Residual Network,IRes Net)customer repayment behavior prediction model.Which is constructed in four steps: Firstly,the SENet is fused into the output vectors of the residual blocks of each layer;Secondly,the fused output vectors are merged into the feature fusion layer;Then,the feature vectors in the feature fusion layer are Global Average Pooled;Finally,the processed feature vectors are input to the Soft Max classification layer to obtain the classification results of customer repayment behaviors.The IRes Net prediction model can not only integrate shallow features with deep features to obtain multi-level information,but also adaptively adjust the spatial feature response values of each channel by virtue of the SENet’s idea of the attention mechanism,thus enhancing the expression capability of the feature vectors in both the level dimension and the spatial dimension,and ultimately making the model have stronger prediction performance.(2)To address the problem of poor model prediction performance caused by both sample imbalance in online lending data and the limitations of single models,based on the IRes Net model,this study further proposed a new customer repayment behavior prediction model(IRes Net-FL-RF)that integrates the Focal Loss and Random Forest.The IRes Net-FL-RF model,uses the Focal Loss function to solve the problems caused by sample imbalance,in addition,considering the advantages of the Random Forest in classification prediction,this research filled the shortcomings of single models by combining of model fusion.Ultimately,the predictive performance of the customer repayment behavior prediction model is further enhanced.(3)In the end,this research validated the prediction performance of the two proposed customer repayment behavior prediction models,by conducting numerous comparative experiments,based on a publicly available dataset from the online lending platform Lending Club.The experimental result indicated that the two models proposed in this research have stronger prediction performance compared to four machine learning-based models and two deep learning-based models in predicting customer repayment behaviors.The experimental results of this research have suggested a new approach for reducing operational risk in the online lending business of financial institutions.