Semi-Supervised Feature Selection with Bias Mitigation for SME Credit Assessment Using Alternative Data

Abstract

Credit scoring for small and medium enterprises (SMEs) faces a fundamental challenge: assessing creditworthiness when traditional financial data is unavailable. This paper presents a semi-supervised feature selection framework that addresses this challenge by leveraging alternative data sources, ranging from transaction patterns to behavioral signals. We develop a graph-based approach that reduces the requirement for labeled data by 70% while improving the area under the curve (AUC) from 0.836 to 0.871, a 4.2 percentage point increase (~5% relative) compared to the best supervised baseline. The framework integrates bias mitigation techniques, which reduce the approval-rate gap by 78.9% while maintaining stable default rates across groups, without compromising model performance. Experiments on 111,579 SME loan applications across three geographic regions demonstrate that the approach scales efficiently with O (n log n) complexity and can process 500,000 applications in approximately two hours (≈131 minutes for 500k records). The practical implications are significant: financial institutions can now assess credit risk for businesses previously considered "unscorable" due to the absence of traditional credit history. This framework facilitates broader access to capital for millions of SMEs, particularly in developing economies where formal financial records are limited.