Software defect prediction based on counterfactual explanations

Abstract

Software defect prediction is critical to ensure the software quality. Researchers have worked on building various defect prediction models to improve the performance of defect prediction. The existing defect prediction models are mainly divided into two categories: one is the models constructed based on the artificial statistical features and the other is the models constructed based on the semantic features. DP-CNN [29] is one of the best defect prediction models, because it combines both the artificial statistical features and semantic features, so its performance is greatly improved compared with traditional defect prediction models. In this paper, we are inspired by the DP-CNN model and make the following two improvements: firstly, we use a new network representation technique Struc2vec to mine the information existing between software modules, which specializes in learning node representations from structural identity and can further extract the structural features associated with defects. We let the DP-CNN model again incorporate the newly mined structural features. Then, we propose a feature selection method based on counterfactual explanations which preserves reasonable causal relationships between features and thus produces a better subset of features compared to traditional feature selection methods. We used it to optimize the artificial statistical features inside the DP-CNN model. Based on the above methods, we propose a new hybrid defect prediction model DPS-CNN-STR. We evaluate our model on six open-source projects in terms of F1-score in defect prediction. The experimental results show that DPS-CNN-STR improves the state-of-the-art method by 3.6% on average.