MVIM-SDM: A Transfer Learning Framework for Credit Classification with Missing Values and Low Sample Size
Abstract
Credit classification in new or restarted business scenarios is often constrained by low sample size in the target domain and missing attribute values in available credit data. This study proposes a novel transfer learning framework titled MVIM-SDM that integrates missing value processing and cross-domain distribution matching into a unified end-to-end learnable architecture for target domain credit risk classification. The proposed Missing Values Ignoring Mask (MVIM) module handles incomplete credit data without conventional imputation by ignoring the contribution of missing values during representation learning, thereby reducing the disturbance introduced by imputed values and improving generalization in low sample size settings. The proposed Spectral Distribution Matching (SDM) module reduces source-target distribution discrepancy in the learned feature space by matching label-consistent spectral representations in the frequency domain, which improves cross-domain knowledge transfer under limited target domain supervision. Experiments on two real-world credit datasets show that MVIM-SDM outperforms the non-transfer learning baseline, domain adaptation methods, feature alignment methods and missing value processing methods, while the ablation results indicate that both modules contribute to the framework. This implies that the proposed framework provides an effective solution to credit classification when target domain data are scarce and incomplete.