Model Transfer and Small-Sample Adaptation for Financial Risk Assessment Across Heterogeneous Domains via Transfer Learning

In practical machine learning applications, achieving robust model transfer across heterogeneous domains with limited labeled data remains a challenging open problem. This paper introduces a novel transfer learning framework that explicitly combines structural risk minimization and kernel-based domain adaptation in a unified model, offering theoretical guarantees for generalization and robustness even under severe data scarcity and domain heterogeneity. The approach projects both source and target data into a high-dimensional feature space — known as a Reproducing Kernel Hilbert Space (RKHS) — where it jointly optimizes predictive accuracy, model simplicity, and alignment between domain distributions. Empirical results on three public credit risk datasets show that the proposed method outperforms strong baselines, improving classification accuracy by 4–7% and achieving consistent gains in F1-score and AUC under small-sample target conditions. These findings demonstrate the method’s superior performance and robustness in real-world scenarios such as cross-domain risk assessment and anomaly detection, and its applicability to domains including finance, healthcare and sensor networks.