Heuristic strategies for solving the combinatorial optimization problem in real-world credit risk assessment

Yongfeng Gu, Hao Ding, Kecai Gu, Runsheng Gan, Xiaoguang Huang, Yanming Fang, Zhigang Hua, Hua Wu, Jifeng Xuan, Jun Zhou

Abstract: Credit risk assessment plays an important role in financial institutions and banks. A common method of predicting credit risks is to build an interpretable rule-set model to distinguish high-risk and non-risk users. This further assists experts and analysts to make critical decisions. In a rule-set model, the quality of selected rules dramatically influences the prediction accuracy. It is challenging to automatically select rules for a rule-set learning task. The rule selection problem in credit risk assessment is considered as a combinatorial optimization problem: the goal is to find a subset of rules that maximize the coverage of risky users under given constraints. In this paper, we employed four methods on top of traditional heuristic strategies, including the Simulated Annealing (SA), Beam Search (BS), Roll-In Roll-Out (RIRO), and Dynamic-step Search (DS). Experiments on real-world datasets show that the proposed heuristics based on SA, BS, and RIRO can detect the most high-risk users in different datasets.