摘要：Software defect prediction enablessoftware developers to estimate the most defective code parts in order to reduce testing efforts. As the size of software project becomes larger, software defect prediction becomes an urgent need. While static product metrics have been extensively investigated as a static meansto predict software defects, coverage analysis of the software has been abandoned due to the expected complexities. Thispaper proposedanovelhybrid approach that leverages code coverage metrics to improve software defect prediction. We build and compare software defect prediction results for four distinct scenarios:static product, code coverage, hybrid, and feature selection. First scenario resembles static analysis and acts as baseline model. Second scenario addresses coverage issues of the associated test cases for the source code. Third and fourth scenarios are derived from combinations of static product and code coverage scenarios. Each scenario has been modeled and examined using thirteen different machine learning classifiers. Two rounds of experiments have been done. First round employs real data extracted from 23 successivereleasesof Apache Lucene, whereas second round applies oversampling technique for the same releases. The results indicate that code coverage scenario attains a significant improvement in software defect prediction,especially when there is a high-coverage ratio for software modules. In general, hybrid scenario outperforms the other three scenarios. Naive Bayesclassifier attains the best results among all classifiers at the first round, while IBK performs wellfor the second round. The second round experiment exhibits a superior performance compared to the first roundbecause itapproaches two times better recall. Further, we notice a steady improvement in the latest releases of Apache Lucene project compared to the earlier ones.
