(A) novel approach shifting the applicable timing of cross-project defect prediction to implementation phase

Abstract

Software Defect Prediction (SDP) is an actively researched technique in Software Engineering (SE). It aims to predict the ongoing software's defect-prone modules (e.g., files, classes) using a prediction model trained with historical defect data. The prediction results through SDP can help allocate Software Quality Assurance (SQA) resources effectively to the defect-prone modules. Among the various types of SDP, Cross-Project Defect Prediction (CPDP) has been used when it is not feasible to collect historical defect data from the ongoing project. It predicts defect-prone modules in an ongoing project (target project) using the prediction model trained with defect data from previously completed projects (source project). However, CPDP encounters a challenge regarding its poor prediction performance, mainly due to differences in the data distribution between the source and target projects. To mitigate this challenge, Transfer Learning (TL) has been employed to reduce the distribution discrepancies between them before constructing a prediction model. Even though TL can improve the prediction performance of CPDP, it requires the target project’s data distribution, which can be obtained only after all modules of the target project are completed (i.e., After implementation phase). Consequently, TL-based CPDP cannot be used to predict the defect-proneness of individual modules when they are completed (i.e., Implementation phase). The infeasibility of TL-based CPDP in the implementation phase hinders the opportunity to ensure the reliability of ongoing software at a lower cost. In addition to this, the TL-based CPDP is inapplicable in various iterative development processes, such as Agile. To address the abovementioned issues, we propose a novel framework feasible in the implementation phase, which is eCPDP (early CPDP), applicable in an earlier phase than exisiting TL-based CPDP techniques. We took advantage of Singular Value Decomposition (SVD) to overcome the reliance on the target project’s data distribution of existing TL techniques. The SVD-based TL method extracts latent factors from the source project data and aligns the target project's individual module data with the extracted latent factors, thereby reducing the distribution discrepancy between source and target project data. To validate the prediction performance of eCPDP, we conducted statistical significance and effect size tests, comparing it with 8 state-of-the-art CPDP techniques across 24 projects and employing 7 evaluation metrics. Additionally, we analyzed the effectiveness of quality assurance resource allocation based on defect prediction results and the overall costs incurred by applying the defect prediction models, considering the different application timing compared to existing techniques. The analysis results indicate that eCPDP outperforms the 8 state-of-the-art CPDP techniques in all 7 metrics in more than half of the subject projects (12 projects), even without the target project’s data distribution. eCPDP enhanced the effectiveness of quality assurance resource allocation by 28% compared to existing techniques and reduced the total cost of applying a defect prediction model by 17%. In conclusion, by applying eCPDP in the construction phase, we anticipate an earlier enhancement in software reliability across various development processes and reduced software quality costs.