Empirical Evaluation of Machine Learning Classifiers for Static Metric Software Defect Prediction

Abstract

With the exponential scaling and structural dependency of complex code deployment, maintaining high software quality parameters has surfaced as a primary bottleneck within the contemporary software development lifecycle. Late-stage bug debugging and software defects noticeably amplify overall implementation costs, consume vast development schedules, and exhaust institutional engineering hours. Consequently, early fault localization and predictive defect classification act as key economic levers, enabling optimal programmatic planning and allocation of limited validation resources. This paper provides a result-oriented comparative mapping of data mining schemes optimized specifically for empirical Software Defect Prediction (SDP) using supervised classification rules. We comprehensively detail the theoretical profiles of multiple paradigms: Bayesian modelling, rule-based induction engines, logistic formulations, and decision tree topologies. Utilizing statistical hold-out validation across twelve high-dimensional legacy systems extracted from the public NASA Metric Data Program (MDP) repository, we baseline the performance of eight prominent learners—Naive Bayes (NB), Logistic Regression (LOG), Decision Table (DT), OneR, PART, JRip, J48, and J48Graft. Evaluation across standard predictive dimensions (Accuracy, Sensitivity, Specificity, Balance, and Receiver Operating Characteristic (ROC) Area) proves that model predictive dominance varies considerably by system architectural traits. The findings confirm that optimizing a defect framework demands a modular, context-driven selection strategy rather than a static monolithic algorithm.