Credit Default Prediction Model based on Horizontal Federated Neural Network and Improved TrAdaBoost Algorithm

Credit default prediction has long been the focus of researchers, with machine learning reshaping its future in unprecedented ways. Due to the limitations of traditional machine learning algorithms in cross-institution data sharing and knowledge transfer, federated learning and transfer learning are gradually gaining prominence in the field of credit default prediction. However, federated learning and transfer learning methods in the field of credit default prediction have the following limitations: (1) The lack of a simple and effective federated learning incentive mechanism. (2) Transfer learning methods lack robustness, with few models capable of handling both similar and dissimilar feature distributions simultaneously. (3) Lack of models that integrate both federated learning and transfer learning methods. In this paper, we present a credit default prediction model that combines a Horizontal Federated Neural Network (HFNN) based on incentive mechanism with an improved TrAdaBoost algorithm. We primarily addressed the following issues: (1) Constructing a federated neural network to handle credit default prediction and introducing an incentive mechanism based on client computational power and local model accuracy. This improved model accuracy and encouraged clients to invest resources in training. (2) Combining federated learning with instance transfer methods to solve data sharing issues and generalize the model's applicability. (3) Improving the TrAdaBoost algorithm by introducing feature alignment and ensemble clustering methods, significantly enhancing its effectiveness in using data with different feature distributions. Finally, we apply the trained model to different credit default prediction scenarios and achieve good performance on the test sets.