On Fault Localization using Machine Learning Techniques

Abstract

Software fault prediction is helpful in early detection of faults, in reducing testing cost and in improving resulting software quality. Software fault prediction considering extracted data from previously tested applications with faults proneness helps to predict software faults in system under test (SUT). Such data sets are collected considering established software metrics proposals such as CK, McCabe and MOOD and are available both as public repositories such as PROMISE1 and as private repositories. We use CK metrics for software fault prediction and we do class-level fault localization, considering dominant features providing us better results, so that we are able to write test cases covering specific aspects of code. We use classification-based software fault predication to arrive at a point where we predict module level faulty parts of SUT and we do fault localization using clustering selecting features using Principal Component Analysis (PCA). We make use of CK metrics and we consider more than sixty data sets having labeled data. We provide validation of our results considering case studies where we first report faulty modules leading to individual classes and considering complete CK metrics producing confusion matrix and comparing results.