Software fault prediction based on improved fuzzy clustering

Abstract

Predicting parts of the software programs that are more defects prone could ease up the software testing process and helps effectively to reduce the cost and time of developments. Although many machine-learning and statistical techniques have been proposed widely for defining fault prone modules in software fault prediction, but this area have yet to be explored with high accuracy and less error. Unfortunately, several earlier methods including artificial neural networks and its variants that have been used, marred by limitations such as inability to adequately handle uncertainties in software measurement data which leads to low accuracy, instability and inconsistency in prediction. In this paper, first the effect of irrelevant and inconsistent modules on fault prediction is decreased by designing a new framework, in which the entire project's modules are clustered. The generated output is then passed to the next model in the hybrid setting, which is a probabilistic neural network (PNN) for training and prediction. We used four NASA data sets to evaluate our results. Performance evaluation in terms of false positive rate, false negative rate, and overall error are calculated and showed 30% to 60% improvement in false negative rate compared to other well-performed training methods such as naïve Bayes and random forest.