Performance Analysis of Feature Selection Techniques in Software Defect Prediction using Machine Learning

Abstract

Software Testing is an essential activity in the development process of a software product. A defect-free software is the need of the hour. Identifying the defects as early as possible is critical to avoid any disastrous consequences in the later stages of development. Software Defect Prediction (SDP) is a process of early identification of defect-prone modules. Lately, software defect prediction coupled with machine learning techniques has gained momentum as it significantly brings down maintenance costs. Feature selection (FS) plays a very significant role in a defect prediction model's efficiency; hence, choosing a suitable FS method is challenging when building a defect prediction model. This paper evaluates six filter-based FS techniques, four wrapper-based FS techniques, and two embedded FS techniques using four supervised learning classifiers over six NASA datasets from the PROMISE repository. The experimental results strengthened that FS techniques significantly improve the model's predictive performance. From our experimental data, we concluded that SVM based defect prediction model showed the best performance among all other studied models. We also observed that Fisher's score, a filter-based FS technique, outperformed all other FS techniques studied in this work.