Software Defect Prediction via GIN with Hybrid Graphical Features

Abstract

Software defect prediction (SDP) is one of the determining factors in software development cycle, enabling developers to estimate the quality of delivery and thus improve the user experience. Generally, efforts have opted for building the abstract syntax tree (AST). However, AST is limited in its capacity of providing adequate semantic and structural information. This paper proposes a framework, which trains hybrid features combined from AST and control flow graph (CFG) via Graph Isomorphism Network (GIN) (H-GIN): The framework parses the CFG of the source code and converts redundant blocks to a concise representation by extracting AST. In addition, the contextual relationships between each block in the CFG is carefully preserved, after which the graphical representation of the code is obtained and then fed into the GIN to conduct a defect prediction. The experiments were performed on a large python dataset called PyTraceBus with Precision, Recall, and F1-score as evaluation criteria. Our approach outperforms the previous state-of-the-art methods, which clearly demonstrates the effectiveness of the proposed method.