Cross-Project Change-Proneness Prediction with Selected Source Project

Abstract

Software change-proneness prediction aims to identity change prone parts of a software where focused attention is required by the managers and other stakeholders. This reduces development and maintenance costs by highlighting the classes which may change and work with the class in such a manner that prevents further changes from occurring often. Prediction requires training data which is generally obtained from historical data of the projects. However, this may not be the case for new projects which have limited or no historical data available. Cross-project change prediction helps solve this issue by using another project as training data to create a prediction model. With the vast number of candidate projects that can be used as a source to train the classifier, the problem of how to select an appropriate source project which can return a decent prediction accuracy with a model trained with it arises in cross-project change prediction.Through this paper, we propose an algorithm to select a source project which can be used to determine change prone classes in a target project with high accuracy. The source project is selected from a pool of 8 open-source projects. Three strategies are used to identity a suitable source project. The results of the three strategies are compared with one another and with a related change-proneness model proposed by Malhotra and Bansal known as the Random Cross-Project Prediction (RCP). Out of the three strategies in the proposed algorithm, the first two strategies performed better in comparison to the prediction performance of the random cross project prediction model with improvements in terms of AUC (1.04% and 1.27%), F-Measure (5.83% and 3.82%), and MCC (14.14% and 7.77%).