Just-in-time software defect prediction based on feature selection

Abstract

At the present stage, just-time software defect prediction has garnered significant attention from researchers due to its granularity and immediacy. Primarily utilizing machine learning classifiers, these models are trained on information from code repositories to predict whether future changes may lead to defects. However, a current challenge with these classifiers lies in the vast number of features, leading to decreased prediction efficiency. These features not only impact model performance but can sometimes result in a decline in predictive accuracy. This paper explores a feature selection technique that combines random forests and self-attention to discard less important features without compromising performance. Through this approach, the number of features required for training is significantly reduced, often to less than 50% of the original features. In our study across six software projects, we observed that using feature selection in the KNN model led to a 9% improvement in the F1 metric and a 6% improvement in the AUC metric compared to logistic regression and Bayesian models. Finally, we applied SHAP for interpretability analysis of the model. This research contributes to enhancing the accuracy and efficiency of just-in-time software defect prediction, providing valuable insights for research and practice in related fields.