Comprehensive Feature Extraction for Cross-Project Software Defect Prediction

Abstract

Quality of software is determined by analyzing the source code metrics and defects. Prediction of pre-and post-release defects and fixing them when they are raised will improve overall quality of software systems. The main aim of this work is to reduce the cost incurred in identifying defects which is a challenging task. Over a decade, the academia and industry addressed the problem of software defect prediction using machine learning. This work emphasizes on the selection of static code metrics and process metrics for defect prediction. Metrics are mined from different open source projects hosted on GitHub. For our experiments, we have chosen projects with at least 10k commits. This paper is establishing a hypothesis that, as the number of commits increase, the bugs are also likely to increase, as our experiments indicate. The software metrics are carefully chosen to identify defects across software projects to help the development and testing teams.