Selection of Significant Metrics for Improving the Performance of Change-Proneness Modules

Abstract

Change-proneness modules are described as the programming parts in the source code which has a high likelihood to modify later on. Change-proneness prediction causes programming analyzers to streamline and focus their testing resources on the modules which have a higher likelihood of modification. Accurate estimation of characteristics such as effort, quality, and risk which are the major concerns for change proneness, is of significant worry in the software life cycle. Regression analysis and the neural network techniques are the commonly used methods for attribute estimation, as per the literature. Chidamber and Kemerer metric (CKJM) suite have been used in this study as an input for model training using neural network with various algorithms to optimal weights and ensemble techniques. These models are validated using five different version of eBay web services. The efficiency of the developed models are computed using three different performance parameters such as AUC, F-Measure, and Accuracy. The information present in the experimental results suggested that the models trained using levenberg marquardt (LM) method achieved better results when compared to the model built using the other classifiers.