From Table to Image: Boosting Credit Risk Prediction via Transfer MLP-like Network on Structured Data

Abstract

At present, deep learning has limited application in the field of financial credit risk because deep learning is good at processing unstructured data such as images, voice, and text, while the credit risk field processes structured tabular data, which makes the existing deep learning methods not well adapted to financial structured data tasks. To this end, this paper proposes a new Table-to-Image Converted Transfer MLP-like network for financial credit risk prediction. First, our method attempts to represent structured data from a new perspective and proposes a data homology based table-to-image conversion method to convert the tabular financial credit risk prediction data into image-like financial data. Then, based on the Strip-MLP structure, a pretrained MLP-like network is proposed to be applied to the credit prediction of the converted image-like financial data. The model is pre-trained with a public financial dataset, and its pre-trained parameters are transferred to the private dataset of financial institutions with different feature numbers and feature contents through transfer learning. Experimental results show that for the task of financial credit risk prediction, the methods proposed in this paper have significantly improved the effect compared with the baseline algorithm.