RASF-LCS: Ranked Attribute Selection and Distance-Based Rule Filtering for Interpretable Credit Scoring

Abstract

Machine learning is widely used in credit scoring, but many high-performing models lack interpretability, limiting their use in regulated domains. Rule-based approaches such as Learning Classifier Systems (LCS) offer a balance between accuracy and explainability. This paper introduces Ranked Attribute Selection with Midpoint Filtering (RASF), an extension to LCS that enhances feature selection and rule validation. RASF combines mutual information-based feature ranking, guided attribute selection, and distance-based rule filtering. Experiments on loan approval datasets show that RASF improves accuracy by 3–5% over standard LCS while maintaining interpretable, rule-based outputs. These results highlight the potential of RASF-LCS for explainable credit decision-making.