Predicting Software Errors Using Feature Selection and Machine Learning Algorithms: A Comparative Study

Software defect prediction plays a key role in improving software quality and optimally allocating testing resources. However, the presence of irrelevant and redundant features in software datasets can increase model complexity and reduce prediction accuracy. In this paper, the impact of ten feature selection methods from three filtering, masking, and embedding approaches on the performance of six machine learning classifiers, including logistic regression, random forest, support vector machine, k-nearest neighbor, naive Bayes, and XGBoost, is investigated through a comparative experimental study. The experiments are conducted on four datasets (CM, KC, KCY, and PC) from the PROMISE repository. The results show that reducing the feature space not only does not reduce prediction accuracy, but in many cases improves it. The combination of mutual information and random forest-based feature importance provided the most stable results, achieving an average F score of ۰.۸۵ when combined with XGBoost, which is an improvement of ۹% over using all features. Furthermore, the selected feature subsets yielded an average dimensionality reduction of ۵۰%. Statistical validation with the Friedman test and Nemenyi post hoc analysis at a significance level of ۰.۰۵ confirms the accuracy of these findings. The findings of this study provide practical guidance for researchers and practitioners who develop machine learning models for software defect prediction.

کلیدواژه ها:

software defect prediction ، feature selection ، machine learning ، static code metrics ، dimensionality reduction ، XGBoost