基于SSDBSCAN的跨项目缺陷预测数据筛选方法

摘要

Since massive irrelevant cross-project data degrades the performance of cross-project defect prediction model in cross-project defect prediction, a data filtering method based on SSDBSCAN (i.e., SSDBSCAN filter) was proposed for cross-project defect prediction.Firstly, the method combined limited amount of labeled within-project data, cross-project data and unlabeled within-project data.Secondly, sub-clusters were found by using SSDBSCAN algorithm.Finally, cross-project data in the sub-clusters was collected and some cross-project data which did not belong to any sub-clusters was discarded as noisy data.The experiments were conducted on 15 public PROMISE datasets using three different classifiers with four performance metrics.Compared with Burak Filter and DBSCAN Filter, the experimental results showed that SSDBSCAN Filter increased PD value, reduced PF value and achieved higher G-measure and AUC values.%针对跨项目软件缺陷预测中大量不相关的跨项目数据损害了缺陷预测模型性能的问题,提出了一种基于SSDBSCAN(semi-suppervised density-based clustering)的跨项目缺陷预测数据筛选方法——SSDBSCAN filter.首先,SSDBSCAN filter结合少量带类标号的本项目历史数据、跨项目历史数据和大量不带类标号的本项目数据;然后,利用SSDBSCAN算法对这些数据进行聚类发现子簇;最后,收集子簇中的跨项目数据,不属于任何簇的跨项目数据被作为噪声数据而丢弃.实验使用15个公开的PROMISE数据集,3种分类器和4种性能度量指标.实验结果表明,相比于目前已有的Burak filter和DBSCAN filter方法,SSDBSCAN filter在提高了预测率的同时降低了误报率,且G-measure与AUC度量值更佳.