A Study of Filter-Based Feature Selection in Software Fault Prediction

Abstract

Software fault prediction (SFP) assists developers in diagnosing the potential defects in the early stage. In SFP, software metrics have strong influence on the performance of a predictive model. However, high dimensional data impacts negatively on the predictive accuracy. As a solution, feature selection provides a process of selecting the optimal features that combine with machine learning techniques to build SFP models. For feature selection, filter selection is a way of addressing the high dimensionality, reducing computation time and improving prediction performance. In this research, we investigate a comparative analysis to review how different of nine filter feature selection methods on both datasets in PROMISE repository, namely CM1 and KC1. The experimental results show that the performances of classifiers are varying on different datasets, especially, in the CM1 dataset, Gain Ratio and Relief based on XGBoost (XGB) and Extra Trees (ET) achieved the highest accuracy and AUC values. In KC1, Gain Ratio and Mutual Information presented the greatest performance among nine methods.