Predicting software fault proneness model using neural network

Abstract

Importance of construction of models for predicting software quality attributes is increasing leading to usage of artificial intelligence techniques such as Artificial Neural Network (ANN). The goal of this paper is to empirically compare traditional strategies such as Logistic Regression (LR) and ANN to assess software quality. The study used data collected from public domain NASA data set. We find the effect of software metrics on fault proneness. The fault proneness models were predicted using LR regression and ANN methods. The performance of the two methods was compared by Receiver Operating Characteristic (ROC) analysis. The areas under the ROC curves are 0.78 and 0.745 for the LR and ANN model, respectively. The predicted model shows that software metrics are related to fault proneness. The models predict faulty classes with more than 70 percent accuracy. The study showed that ANN method can also be used in constructing software quality models and more similar studies should further investigate the issue. Based on these results, it is reasonable to claim that such a model could help for planning and executing testing by focusing resources on fault-prone parts of the design and code. © 2008 Springer-Verlag Berlin Heidelberg.