Feature Selection for Machine Learning Fault Prediction in Software Modules

Abstract

Machine learning-based fault prediction depends strongly on the quality of software metrics used as input features. Large module-level metric sets often contain redundant, weak, noisy, or highly correlated variables that can reduce prediction clarity and increase model complexity. This article presents a feature selection framework for software module fault prediction using correlation screening, Information Gain ranking, Chi-Square ranking, wrapper-based subset evaluation, and combined filter-wrapper refinement. The framework evaluates software metrics related to code size, complexity, coupling, churn, developer activity, prior defects, and dependency behavior before selecting compact feature subsets for classification. The study shows that feature selection improves fault prediction by reducing unnecessary metrics, strengthening faulty-module recall, lowering false positives, and improving the interpretability of selected risk indicators. The proposed approach supports more focused defect-prone module identification and helps software quality teams prioritize inspection using a smaller, more meaningful metric set.