A Model Based on Program Slice and Deep Learning for Software Defect Prediction

Abstract

Defects are inherent in software and can lead to many serious problems during the use of software. Software defect prediction is an important method for finding defects and can help developers improve their testing efficiency. To build accurate prediction models, previous software defect prediction techniques focus on design of functions related to the code of potential defects. But these methods do not adequately capture the semantic features of the program. In this paper, we propose a software defect prediction model based on program slice and deep learning. We extract program slice based on system dependence graph and leverage Gated Recurrent Unit (GRU) to generate features. We conducted experiments both within-project defect prediction and cross-project defect prediction using the dataset of the public PROMISE database. The results show that our method improves on average by 11.0% in F1-measure in within-project and 10.4% in F1-measure in cross-project defect prediction compared to the Tree-LSTM method.