Meta network attention-based feature matching for heterogeneous defect prediction

Abstract

Cross-project defect prediction (CPDP) involves predicting defects in projects without historical data by utilizing information from other projects. This requires uniform metrics across source and target projects (CPDP-CM). However, heterogeneous defect prediction (HDP), which deals with different metric sets, faces challenges such as feature alignment and distribution inequalities. This paper addresses these challenges with a novel method: Meta Network Attention-based Feature Matching (MNAFM) for HDP. Our approach uses a meta-network to identify feature similarities and adjust distillation intensity, enhancing HDP accuracy. Experiments on 30 projects demonstrate that MNAFM significantly outperforms baseline methods, showing improvements in f-measure (11.42–64.12%), g-measure (11.94–30.12%), and MCC (16.58–98.63%). Statistical tests confirm that MNAFM outperforms eight benchmark algorithms. Additionally, an ablation study highlights the contribution of each component of MNAFM, demonstrating the importance of the attention mechanism, data augmentation, symmetrical padding, and the use of a pretrained ResNet model in achieving superior performance. In summary, MNAFM offers a significant advancement in heterogeneous defect prediction by effectively leveraging feature similarities, distillation adjustments, and a robust methodological framework.