Towards a Neural Network based Reliability Prediction Model via Bugs and Changes

Nowadays, software systems have become larger and more complex than ever. A system failure could threaten the safety of human life. Discovering the bugs as soon as possible during the software development and investigating the effect of a change in the software system are two main concerns of the software developers to increase system's reliability. Our approach employs a neural network to predict reliability via post-release defects and changes applied during the software development life cycle. The CK metrics are used as predictors variables, whereas the target variable is composed of both bugs and changes having different weights. This paper empirically investigates various prediction models considering different weights for the components of the target variable using five open source projects. Two major perspectives are explored: cross-project to identify the optimum weight values for bugs and changes and cross-project to discover the best training project for a selected weight. The results show that for both cross-project experiments, the best accuracy is obtained for the models with the highest weights for bugs (75% bugs and 25% changes) and that the right fitted project to be used as training is the PDE project.