TWAO: Time-Weight-Aware Oversampling Method for Just-in-Time Software Defect Prediction

Just-in-Time Software Defect Prediction aims to predict whether a developer’s commit change will introduce a defect, so that potential problems can be quickly located and fixed. However, software defect data distribution is often highly imbalanced, which will significantly affect the performance of the prediction model. Many researchers have used oversampling methods to address this problem. Still, existing oversampling techniques are usually performed from the perspective of the sample feature space, which cannot be adapted to the temporal nature of Just-in-Time software defect prediction. Therefore, this paper proposes a Time-Weight-Aware Oversampling method (TWAO) for Just-in-Time Software Defect Prediction. TWAO uses temporal information as a guiding factor to make the new samples generated more diverse and the data distribution more reasonable to alleviate the class imbalance problem better. In this paper, we compare the performance of TWAO with 8 oversampling methods on 14 open-source defect datasets and 5 prediction models. The experimental results show that TWAO effectively improves the performance of the prediction models and is an efficient solution for building Just-in-Time Software Defect Prediction models on imbalanced datasets.