Research Article
Heterogeneous Graph Neural Network-Based Software Developer Recommendation
@INPROCEEDINGS{10.1007/978-3-031-24383-7_24, author={Zhixiong Ye and Zhiyong Feng and Jianmao Xiao and Yuqing Gao and Guodong Fan and Huwei Zhang and Shizhan Chen}, title={Heterogeneous Graph Neural Network-Based Software Developer Recommendation}, proceedings={Collaborative Computing: Networking, Applications and Worksharing. 18th EAI International Conference, CollaborateCom 2022, Hangzhou, China, October 15-16, 2022, Proceedings, Part I}, proceedings_a={COLLABORATECOM}, year={2023}, month={1}, keywords={Developer recommendation Heterogeneous graph neural network Information supplement Self-supervised learning}, doi={10.1007/978-3-031-24383-7_24} }- Zhixiong Ye
Zhiyong Feng
Jianmao Xiao
Yuqing Gao
Guodong Fan
Huwei Zhang
Shizhan Chen
Year: 2023
Heterogeneous Graph Neural Network-Based Software Developer Recommendation
COLLABORATECOM
Springer
DOI: 10.1007/978-3-031-24383-7_24
Abstract
In software maintenance, it is critical for project managers to assign software issues to the appropriate developers. However, finding suitable developers is challenging due to the general sparsity and the long-tail of developer-issue interactions. In this paper, we propose a novelHeterogeneousGraph Neural Network-based method forDeveloperRecommendation (called HGDR), in which text information embedding and self-supervised learning (SSL) are incorporated. Specifically, to alleviate the sparsity of developer-issue interactions, we unify developer-issue interactions, developer-source code file interactions and issue-source code file relations into a heterogeneous graph, and we embed text descriptions to graph nodes as information supplements. In addition, to mitigate the long-tail influence, e.g., recommendation bias, the proficiency weight suppression link supplementation is proposed to complement the tail developers by adjusting proficiency weights. Finally, to fully utilize rich structural information of heterogeneous graph, we use the joint learning of metapath-guided heterogeneous graph neural network and SSL to learn the embedding representation. Extensive comparison experiments on three real-world datasets show that HGDR outperforms the state-of-the-art methods by 6.02% to 44.27% on recommended metric. The experimental results also demonstrate the efficacy of HGDR in the sparse and long-tail scenario. Our code is available athttps://github.com/1qweasdzxc/HGDR.
