Assessing Software Quality by Program Clustering and Defect Prediction

Many empirical studies have shown that defect prediction models built on product metrics can be used to assess the quality of software modules. So far, most methods proposed in this direction predict defects by class or file. In this paper, we propose a novel software defect prediction method based on functional clusters of programs to improve the performance, especially the effort-aware performance, of defect prediction. In the method, we use proper-grained and problem-oriented program clusters as the basic units of defect prediction. To evaluate the effectiveness of the method, we conducted an experimental study on Eclipse 3.0. We found that, comparing with class-based models, cluster-based prediction models can significantly improve the recall (from 31.6% to 99.2%) and precision (from 73.8% to 91.6%) of defect prediction. According to the effort-aware evaluation, the effort needed to review code to find half of the total defects can be reduced by 6% if using cluster-based prediction models.