GLA-SDP: A novel attention-based semantic and static feature fusion method using GCN and LSTM for software defect prediction

Software defect prediction (SDP) is crucial for improving software quality and optimizing resource allocation by identifying potential defects at early development stages. Traditional SDP methods rely on manually crafted static features, which often fail to capture the semantic and contextual information in code. Recent advances using abstract syntax trees (ASTs) have improved semantic feature extraction, yet they often neglect structural dependencies and lack effective integration with static features. To address these limitations, this study proposes GLA-SDP, a novel defect prediction model that fuses features through the integration of Graph Convolutional Networks (GCNs), Long Short-Term Memory (LSTM) networks, and an additive attention fusion mechanism. Specifically, a recursive AST-to-graph construction method is designed to extract rich semantic features using GCNs, while LSTMs are employed to capture sequential patterns from static code metrics. Furthermore, an attention-based fusion mechanism dynamically weight and combine semantic and static features, preserving their complementary importance in defect prediction. Extensive experiments on eight Java projects from the PROMISE dataset and four C-language projects from the Devign dataset demonstrate that GLA-SDP consistently outperforms state-of-the-art baselines, achieving average improvements of 37 % in F1-score and 24 % in MCC. These results highlight the superior accuracy and practical applicability of the proposed approach.