CDS: A Cross–Version Software Defect Prediction Model With Data Selection

Over the past decade, a large number of software defect prediction approaches have been proposed to identify the defect-prone modules by mining software repositories. Recently, a novel scenario called Cross-Version Defect Prediction (CVDP) begins to draw increasing research interests, as it is more reasonable and applicable in practice to adopt the labeled defect data of previous versions to predict defects in the current version of the same project. As a software project often has multiple previous versions, CVDP on this kind of projects will face the following two critical but seldom reported issues, namely, data distribution difference and class overlapping. In this paper, we address these two issues by solving a version selection problem via a Cross-version model with Data Selection (CDS). The proposed CDS is a novel framework which treats the defect prediction of existing and new files in different ways. For the existing files, we propose a novel Clustering-based Multi-Version Classifier (CMVC), which can automatically select the training data from the most relevant and noise-free versions by assigning them higher weights than the others. We proposed a Weighted Sampling Model (WSM) for the new files which have never appeared in previous version by incorporating the outputs of CMVC. We evaluate the proposed CDS model on 28 versions across 8 software projects, and the experimental results demonstrate that CDS outperforms three baseline methods and a state-of-the-art approach in terms of three prevalent performance indicators.