Using Clustering And Voting To Generate Fault-prone Module Ranking

The presence of software faults is a huge risk to the safe use of the software. The prediction of fault-prone module ranking, as an effective mean for software quality control, can help software engineer to detect software faults efficiently.However, existing fault prediction models have many limitations when applied in practice. First, supervised-based fault prediction models are not appropriate for newly developed software due to the unavailability of historical fault information in the process of modeling. Second, existing unsupervised-based fault prediction models are mainly for classification rather than ranking. Therefore, they cannot be used to effectively allocate testing resources. Third, semi-supervised-based fault prediction models have a high cost because experts are needed during the modeling process. In particular, most of existing fault prediction models are evaluated by the indicators based on the confusion matrix. These indictors are effort-unaware, i.e., they do not take into account the effort to inspect/test when inspecting/testing these modules that are predicted as faulty. Consequently, it may not be cost-effective when applying these modules in practice.This thesis makes the following three contributions. First, we propose four inter-/inner-cluster ranking methods used for clustering-based fault prediction models. Based on five open-source systems, we use an effort-aware indicator to compare the effectiveness of these four ranking methods. Second, we apply a voting approach to generate a fault-prone module ranking. In particular, we give the principles to choice the voters and vote strategies in the modeling process. The experimental results show that the voting model is more cost-effective than the baseline model that ranks modules in ascending according to their LOC. Third, we combine the clustering-based fault-proneness ranking model with the voting method to generate fault-prone module ranking. The experimental results show that this combination can produce more cost-effective fault-prone module ranking.The main contributions of this thesis are summarized as follows:(1) We proposed four inter-/inner-cluster ranking algorithms used in clustering-based fault-proneness ranking model and experimentally compared their ability in ranking fault-prone modules;(2) We applied the voting method to a number of commonly used structural metrics to produce fault-prone ranking for modules.(3) We use the voting method to improve the clustering-combined module fault-proneness ranking model.