A Learning-to-Rank Algorithm for Constructing Defect Prediction Models

This paper applies the learning-to-rank approach to software defect prediction. Ranking software modules in order of defect-proneness is important to ensure that testing resources are allocated efficiently. However, prediction models that are optimized for predicting explicitly the number of defects often fail to correctly predict rankings based on those defect numbers. We show in this paper that the model construction methods, which include the ranking performance measure in the objective function, perform better in predicting defect-proneness rankings of multiple modules. We present the experimental results, in which our method is compared against three other methods from the literature, using five publicly available data sets.