Nonlinear Discriminant Analysis for Predicting Fault-Prone Modules in Legacy Software

Shuji Takabayashi (9851059)


Legacy software is software that has been developed many years ago and has been continuously modified and expanded till today. However, continuous modification and expansion produces fault-prone modules, that will increase the maintenance costs. In order to lessen such maintenance costs, we need to predict the fault-prone modules in advance, and to test them thoroughly or sometimes redesign them as new modules.

This paper proposes a discriminant analysis method that uses a neural network model to predict the fault-prone modules. Since the relation between fault-proneness and predictor variables are complicated in real software, conventional linear discriminant model is not suitable for prediction model. Therefore, this thesis uses a layered neural network to represent nonlinear relation between fault-proneness and predictor variables.

To evaluate the method, I have measured many kinds of metrics from large-scale software that have been maintained more than 20 years, and also measured the number of faults found after the release. Result of the evaluation showed that prediction accuracy of my model is better than that of conventional linear model.