Analyzing fault prediction usefulness from cost perspective using source code metrics

Abstract

Software fault prediction techniques are useful for the purpose of optimizing test resource allocation. Software fault prediction based on source code metrics and machine learning models consists of using static program features as input predictors to estimate the fault proneness of a class or module. We conduct a comparison of five machine learning algorithms on their fault prediction performance based on experiments on 56 open source projects. Several researchers have argued on the application of software engineering economics and testing cost for the purpose of evaluating a software quality assurance activity. We evaluate the performance and usefulness of fault prediction models within the context of a cost evaluation framework and present the results of our experiments. We propose a novel approach using decision trees to predict the usefulness of fault prediction based on distributional characteristics of source code metrics by fusing information from the output of the fault prediction usefulness using cost evaluation framework and distributional source code metrics.