A Systematic Study of Cross-Project Defect Prediction With Meta-Learning

The prediction of defects in a target project based on data from external projects is called Cross-Project Defect Prediction (CPDP). Several methods have been proposed to improve the predictive performance of CPDP models. However, there is a lack of comparison among state-of-the-art methods. Moreover, previous work has shown that the most suitable method for a project can vary according to the project being predicted. This makes the choice of which method to use difficult. We provide an extensive experimental comparison of 31 CPDP methods derived from state-of-the-art approaches, applied to 47 versions of 15 open source software projects. Four methods stood out as presenting the best performances across datasets. However, the most suitable among these methods still varies according to the project being predicted. Therefore, we propose and evaluate a meta-learning solution designed to automatically select and recommend the most suitable CPDP method for a project. Our results show that the meta-learning solution is able to learn from previous experiences and recommend suitable methods dynamically. When compared to the base methods, however, the proposed solution presented minor difference of performance. These results provide valuable knowledge about the possibilities and limitations of a meta-learning solution applied for CPDP.