The Impact of Feature Selection on Predicting the Number of Bugs

Abstract: Bug prediction is the process of training a machine learning model on software metrics and fault information to predict bugs in software entities. While feature selection is an important step in building a robust prediction model, there is insufficient evidence about its impact on predicting the number of bugs in software systems. We study the impact of both correlation-based feature selection (CFS) filter methods and wrapper feature selection methods on five widely-used prediction models and demonstrate how these models perform with or without feature selection to predict the number of bugs in five different open source Java software systems. Our results show that wrappers outperform the CFS filter; they improve prediction accuracy by up to 33% while eliminating more than half of the features. We also observe that though the same feature selection method chooses different feature subsets in different projects, this subset always contains a mix of source code and change metrics.