Transfer Learning for Cross-Project Change-Proneness Prediction in Object-Oriented Software Systems

Abstract

Change-prone classes or modules are defined as regions of the source code which are more likely to change as a result of a software development of maintenance activity. Automatic identification of change-prone classes are useful for the software development team as they can focus their testing efforts on areas within the source code which are more likely to change. Several machine learning techniques have been proposed for predicting change-prone classes based on the application of source code metrics as indicators. However, most of the work has focused on within-project training and model building. There are several real word scenario in which sufficient training dataset is not available for model building such as in the case of a new project. Cross-project prediction is an approach which consists of training a model from dataset belonging to one project and testing it on dataset belonging to a different project. Cross-project change-proneness prediction is relatively unexplored. We propose a machine learning based approach for cross-project change-proneness prediction. We conduct experiments on 10 open-source Eclipse plug-ins and demonstrate the effectiveness of our approach. We frame several research questions comparing the performance of within project and cross project prediction and also propose a Genetic Algorithm (GA) based approach for identifying the best set of source code metrics. We conclude that for within project experimental setting, Random Forest (RF) technique results in the best precision. In case of cross-project change-proneness prediction, our analysis reveals that the NDTF ensemble method performs higher than other individual classifiers (such as decision tree and logistic regression) and ensemble methods in the experimental dataset. We conduct a comparison of within-project, cross-project without GA and cross-project with GA and our analysis reveals that cross-project with GA performs best followed by within-project and then cross-project without GA.