A Comparison Analysis of Constraint-Handling Techniques on Rule Selection Problem in Credit Risk Assessment: An Industrial View

The rule selection problem, aiming at selecting a subset of rule candidates, is important in credit risk assessment. In reality, we hope these selected rules can identify as many high-risk users as possible while fulfilling the business constraint simultaneously. In such cases, the rule selection problem can be seen as a typical constrained combinatorial optimization problem and can be solved by evolutionary algorithms (EAs). However, how to deal with the constraint dramatically affects the performance of our EAs since the feasible and infeasible solutions both affect the convergence and direction of the EAs. This paper conducts a comparative study to explore the performance of six constraint-handling techniques on the rule selection problem. Experimental results indicate the importance of the way we handle the constraints: 1) e-Constrained Method finds the most high-risk users and Constrained-Domination Principle obtains the most diversified solutions. 2) dominance-based techniques outperform penalty function-based techniques in our real-world applications.