DeepPull: Deep Learning-Based Approach for Predicting Reopening, Decision, and Lifetime of Pull Requests on GitHub Open-Source Projects

This paper introduces DeepPull, a novel multi-output deep learningbased classification approach designed to predict pull request outcomes in GitHub's pull-based software development model. The primary goal of DeepPull is to provide decision-making support for integrators in open-source software projects, with a particular focus onmitigating the challenges associated with managing high request volumes. DeepPull anticipates the reopening, decision, and lifetime of pull requests at the time of submission. Our method effectively leverages diverse data sources, including tabular and textual data, and incorporates a combination of SMOTE and VAE techniques to handle imbalances in reopening predictions. The evaluation of DeepPull on six well-known programming languages, along with 83 open-source projects, demonstrates significant performance improvements over both a randomized baseline and the existing approach. The approach greatly enhances balanced accuracy by 6.25%, AUC by 7.19%, and TPR by 16.78% in reopening prediction, improves accuracy by 7.71%, precision by 0.56%, recall by 10.96%, and F-measure by 6.27% in decision prediction, and reduces MMAE by 5.73% in lifetime prediction compared to an existing approach.