Software fault prediction using the hybrid feature selection: a study on the bughunter dataset

Tóm tắt

Software defect prediction is essential for improving software quality and reducing testing costs. However, defect datasets typically contain high-dimensional, imbalanced, and noisy features that negatively affect machine learning performance. This study proposes a hybrid feature selection approach that integrates four statistical filters - Correlation-based Feature Selection, Fisher Score, Laplacian Score, and Gain Ratio - with the Whale Optimization Algorithm (WOA). Experiments were conducted on seven BugHunter datasets containing 96 original features, using Random Forest as the predictive model. The hybrid method reduces the feature set to 19–28 features and achieves the best performance among all tested configurations, with an average Accuracy of 0.720, F1-score of 0.692, and AUC of 0.768. The results demonstrate that combining multi-filter feature selection with meta-heuristic optimization effectively removes irrelevant features, enhances model stability, and improves predictive capability. This hybrid approach shows strong potential for practical applications in software engineering.