Recommending developers suitable for Pull Requests (PRs) review holds paramount importance in facilitating evolution and knowledge sharing on GitHub. Current researches primarily focus on modeling developer expertise or developer-PR interactions using historical review data. However, these approaches lack consideration for integrating domain-specific PRs and temporal changes into developer recommendations, which results in sub-optimal performance. To address these problems, we introduce a time-aware developer recommendation approach based on multi-feature fusion, called TFRec. We utilize the BERT model, trained via contrastive learning, to embed textual PR information (e.g., title and description), facilitating efficient semantic learning and generalization improvement. Additionally, we employ high-order propagation and embedding to fuse the semantic and structure features of historical data and capture potential collaborative relationships from developer-PR interactions. Furthermore, a time-aware attention is introduced to encode the temporal features into the propagation and embedding process. Extensive experiments on a dataset of 65k PRs from 10 GitHub open-source projects demonstrate the superior performance of our proposed method compared to state-of-the-art approaches.
原文链接
分享
引用
