Comparison of back propagation training algorithms for software defect prediction

Abstract

The cost of deleting a software bug increases ten times as it is floated onto the next phase of software development lifecycle (SDLC). This makes the task of the project managers difficult and also degrades the quality of the output software product. Software defect prediction (SDP) was proposed as a solution to the problem which could anticipate the defective modules and hence, deal with them in an efficient and effective manner in advance. The adequacy of artificial neural networks (ANNs) to handle the complex nonlinear relationships between the software metrics and the defect data demonstrates their suitability to build the defect prediction models. In this paper, multilayer feed forward back propagation based neural networks were constructed using seven defect datasets from the PROMISE repository. An empirical comparison of Levenberg-Marquardt (LM), Resilient back propagation (RP) and Bayesian Regularization (BR) back propagation training algorithms was performed using statistical measures such as MSE and R2 values and the parameters computed from the confusion matrix. Bayesian based back propagation training method performed better than the LM and RP techniques in terms of minimizing mean square error and type II error and maximizing accuracy, sensitivity and R2 value. An accuracy of more than 90 percent was achieved by BR on all the seven datasets and the best data fit during the regression analysis was shown with a R2 value of 0.96. Overall, it is the context and the criticality of the software project which will aid the project managers to prioritize the performance measures and hence, decide upon the training algorithm to be applied, according to the goals and resources available.