An ensemble learning model with dynamic sampling and feature fusion network for class sparsity in credit risk classification

The prevalent challenge of class sparsity issues in credit risk classification commonly focuses on instance-view solutions, while feature-view solutions are overlooked. For this purpose, this paper designs a dual-view ensemble learning model to tackle class sparsity and its associated traits of overlap, noise, and irrelevance. The model comprises two phases integrated into a recurrent structure. Firstly, an instance-view dynamic sampling method is developed on instance importance estimation to select important instances. Secondly, at the feature view, a feature fusion network is introduced to extract classification features by integrating feature interaction and densely connected structures. In order to form a recurrent structure, the trained network serves as an instance importance estimator in the subsequent epoch. The proposed model is evaluated using four publicly available datasets and six derived datasets, and experimental results demonstrate its excellent performance relative to other benchmarks. This indicates the proposed ensemble model presents an effective and competitive solution for credit risk classification in scenarios with class sparsity.