DGRL-SDP: Dual-View Graph Representation Learning for Software Defect Prediction

Existing software defect prediction (SDP) methods focus exclusively on static semantic features, overlooking dynamic execution information (such as data-flow dependencies and method-call sequences), which limits their ability to capture latent defects. To overcome this limitation, this paper proposes a novel approach called Dual-View Graph Representation Learning for Software Defect Prediction (DGRL-SDP). First, DGRL-SDP enriches existing Abstract Syntax Tree (AST)-based representations by incorporating data-flow and method-flow edges, constructing enhanced graphs that capture runtime dependencies. These graphs are then processed by GraphSAGE to extract defect-prone features, which are subsequently fused with deep semantic representations via a gated mechanism to produce a unified embedding. Experiments conducted on eight Java projects from the PROMISE and GitHub Bug datasets demonstrate that DGRL-SDP achieves superior improvements over competitive baselines in F-measure and MCC. Additional empirical results further confirm the effectiveness of incorporating dynamic dependency information.