A Cross-Project Evaluation of Text-Based Fault-Prone Module Prediction

In the software development, defects affect quality and cost in an adverse way. Therefore, various studies have been proposed defect prediction techniques. Most of current defect prediction approaches use past project data for building prediction models. That is, these approaches are difficult to apply new development projects without past data. In this study, we focus on the cross project prediction that can predict faults of target projects by using other projects. We use 28 versions of 8 projects to conduct experiments of the cross project prediction and intra-project prediction using the fault-prone filtering technique. Fault-prone filtering is a method that predicts faults using tokens from source code modules. Additionally, we try to find an appropriate prediction model in the fault-prone filtering, since there are several ways to calculate probabilities. From the results of experiments, we show that using tokens extracted from all parts of modules is the best way to predict faults and using tokens extracted from code part of modules shows better precision. We also show that the results of the cross project predictions have better recall than the results of the intra-project predictions.