Just-in-time software defect prediction using deep temporal convolutional networks

Software maintenance and evolution can introduce defects in software systems. For this reason, there is a great interest to identify defect prediction and estimation techniques. Recent research proposes just-in-time techniques to predict defective changes just at the commit level allowing the developers to fix the defect when it is introduced. However, the performance of existing just-in-time defect prediction models still requires to be improved. This paper proposes a new approach based on a large feature set containing product and process software metrics extracted from commits of software projects along with their evolution. The approach also introduces a deep temporal convolutional networks variant based on hierarchical attention layers to perform the fault prediction. The proposed approach is evaluated on a large dataset, composed of data gathered from six Java open-source systems. The obtained results show the effectiveness of the proposed approach in timely predicting defect proneness of code components.