Just-in-time defect prediction for Android apps via imbalanced deep learning model

Android mobile apps have played important roles in our daily life and work. To meet new requirements from users, the mobile apps encounter frequent updates, which involves in a large quantity of code commits. Previous studies proposed to apply Just-in-Time (JIT) defect prediction for mobile apps to timely identify whether new code commits can introduce defects into apps, aiming to assure the quality of mobile apps. In general, the number of defective commit instances is much fewer than that of clean ones, in other words, the defect data is class imbalanced. In this work, we propose a novel Imbalanced Deep Learning model, called IDL, to conduct JIT defect prediction task for Android mobile apps. More specifically, we introduce a state-of-the-art cost-sensitive cross-entropy loss function into the deep neural network to learn the high-level feature representation, in which the loss function alleviates the class imbalance issue by taking the prior probability of the two types of classes into account. We conduct experiments on a benchmark defect data consisting of 12 Android mobile apps. The results of rigorous experiments show that our proposed IDL model performs significantly better than 23 comparative imbalanced learning methods in terms of Matthews correlation coefficient performance indicator.