Heterogeneous Defect Prediction through Joint Metric Selection and Matching

Defect prediction is one of the hot topics in software engineering. To relax the restriction on metrics, heterogeneous defect prediction (HDP) arises and aims to conduct the prediction across projects with different metrics. Among existing HDP methods, (1) one type of them construct a common metric space for heterogeneous source and target data by metric matching regardless of removing redundant metrics; (2) the other type of them generally consist of two phases that are conducted individually, i.e., metric selection and matching, which makes the whole process likely suboptimal. To solve these issues, we propose a novel approach Jointly optimizing Metric Selection and Matching (JMSM). Specifically, JMSM employs maximum mean discrepancy to reduce the distribution difference between source and target data while filtering out redundant metrics by introducing the ${l_{2,1}}$ -norm regularization. Experiments on 22 projects from three datasets demonstrate the significant superiority of JMSM over baselines in performance and verify the effectiveness of introducing ${l_{2,1}}$ -norm regularization.