Unleashing the power of dynamic ensembles: a risk- and profit-driven framework for online loan credit risk assessment

Effective credit risk assessment is essential in online lending, enabling lenders to make informed investment decisions. Recently, the dynamic ensemble selection (DES) technique, which adaptively selects an appropriate classifier ensemble for each sample, has shown great promise in this domain. However, traditional DES methods often rely on bootstrap sampling to generate base classifiers, potentially overlooking the intrinsic structure of input data and reducing predictive performance. Additionally, existing profit-oriented credit risk models fail to accommodate the heterogeneous risk preferences of investors. To address these issues, we propose a novel hybrid model, DES-CBCG-MOO, which combines (1) a clustering-based classifier generation (CBCG) module to enhance diversity among base classifiers, and (2) a multi-objective optimization (MOO) component that integrates NSGA-III into the META-DES framework to jointly optimize risk and profitability. To enhance model transparency, we employ SHAP for post-hoc interpretation of predictions. Experiments on a real-world lending dataset demonstrate that DES-CBCG-MOO outperforms the baselines.