Semantic Feature Learning via Dual Sequences for Defect Prediction

Software defect prediction (SDP) can help developers reasonably allocate limited resources for locating bugs and prioritizing their testing efforts. Existing methods often serialize an Abstract Syntax Tree (AST) obtained from the program source code into a token sequence, which is then inputted into the deep learning model to learn the semantic features. However, there are different ASTs with the same token sequence, and it is impossible to distinguish the tree structure of the ASTs only by a token sequence. To solve this problem, this paper proposes a framework called Semantic Feature Learning via Dual Sequences (SFLDS), which can capture the semantic and structural information in the AST for feature generation. Specifically, based on the AST, we select the representative nodes in the AST and convert the program source code into a simplified AST (S-AST). Our method introduces two sequences to represent the semantic and structural information of the S-AST, one is the result of traversing the S-AST node in pre-order, and another is composed of parent nodes. Then each token in the dual sequences is encoded as a numerical vector via mapping and word embedding. Finally, we use a bi-directional long short-term memory (BiLSTM) based neural network to automatically generate semantic features from the dual sequences for SDP. In addition, to leverage the statistical characteristics contained in the handcrafted metrics, we also propose a framework called Defect Prediction via SFLDS (DP-SFLDS) which combines the semantic features generated from SFLDS with handcrafted metrics to perform SDP. In our empirical studies, eight open-source Java projects from the PROMISE repository are chosen as our empirical subjects. Experimental results show that our proposed approach can perform better than several state-of-the-art baseline SDP methods.