Similarity-based and rank-based defect prediction

In this paper, we explore two new approaches for software defect prediction. The similarity-based approach predicts the number of latent defects of a software module from those of modules most similar to it. The rank-based approach uses machine learning models specially trained to predict the ranks of software modules based on their actual number of latent defects. In both approaches, we use technical concerns/functionalities recovered by topic modeling techniques as features to represent software modules. Empirical evaluation with five real software systems shows that the proposed approaches outperform the traditional one and a recently introduced defect prediction method.