Treffer: Software Defect Prediction Based on Multi-Hypergraph Adaptive Learning and Ensemble Learning.

Software defect prediction plays a vital role in improving software quality and reliability. However, existing methods based on graph neural networks and hypergraph neural networks remain limited in modeling complex code dependency relations under multi-view feature settings. To address this limitation, we propose a novel multi-hypergraph neural network method, termed MHGNN-E, which constructs multiple view-specific hypergraphs from three complementary perspectives, including static code metrics, complex network metrics, and network embedding metrics, and learns representations via hypergraph convolution. Furthermore, we introduce an adaptive fusion mechanism with dynamic weighting to capture cross-view dependencies. In addition, to mitigate class imbalance, we employ a Bayesian-optimized ensemble classifier to fuse multiple base learners, which further improves prediction stability and generalization. Experiments on 29 PROMISE datasets show that MHGNN-E consistently improves key metrics, including AUC, F1, and MCC. Specifically, compared with the baselines, MHGNN-E improves AUC by 3.56% to 30.49%, F1 by 12.30% to 26.76%, and MCC by 29.07% to 81.98%. These results consistently support the effectiveness of the proposed multi-hypergraph adaptive learning and ensemble strategies for software defect prediction, and provide insights into hypergraph-based software defect prediction. [ABSTRACT FROM AUTHOR]