GAT-Team: A Team Recommendation Model for Open-Source Software Projects

Qing Qi, Jian CaoSEKE 2023 · OpenReview

Introduction

Open-source software (OSS) projects are fundamentally collaborative, yet most contributor recommendation systems focus on individual expertise rather than team dynamics. This approach overlooks the high "startup costs" associated with forming new teams of strangers. Research indicates that "independent teams"—groups of developers who have previously collaborated and established social ties—are significantly more efficient. Empirical analysis of GitHub data reveals that members of these independent teams maintain a contribution rate 2.57 to 2.60 times higher than the average contributor. GAT-Team is a novel recommendation model designed to identify and suggest these high-performing, socially-connected units to OSS projects.

Methodology

The GAT-Team pipeline consists of two primary stages: Team Discovery and Team Recommendation.

Team Discovery

To identify existing teams, the authors construct a Developer Social Network (DSN) where edges represent "mentions" (e.g., @user) in GitHub issue comments. Unlike simple "following" relationships, @mentions serve as a high-signal proxy for direct social interaction and collaboration. The OSLOM (Order Statistics Local Optimization Method) algorithm is then applied to this network to detect overlapping communities. A group is defined as an "independent team" if its members have worked on multiple projects together and possess strong social connections.

The GAT-Team Model

The core of the system is a three-layer Graph Attention Network (GAT) designed to propagate information across the hierarchical entities of the OSS ecosystem:

  1. Repo-Developer Layer: Propagates information between project repositories and their core developers.
  2. Inter-Developer Layer: Refines developer representations by modeling social connections within the DSN.
  3. Developer-Team Layer: Connects individual developer embeddings to their respective independent teams.
The three-layer Graph Attention Network architecture of the GAT-Team model, illustrating information propagation from repositories to teams.

The final output is processed through a feed-forward network with a Sigmoid activation to predict the suitability of a team for a specific repository.

Experimental Setup and Results

The model was evaluated using a massive dataset crawled from GitHub (January 2015 to June 2020), encompassing 312,934 repositories, 1,384,736 users, and over 1.1 million social edges. Using the OSLOM algorithm, 22,875 independent teams were identified for the study.

Performance was measured using nDCG (Normalized Discounted Cumulative Gain). GAT-Team achieved an nDCG@30 of 0.9017, significantly outperforming traditional baselines.

Performance comparison of GAT-Team against various baselines, including social and interest-based models, across different k values.

Key findings from the evaluation include:

  • Superiority over Baselines: GAT-Team vastly outperformed interest-based models (which scored between 0.01 and 0.06) and the Social baseline (0.7376).
  • Failure of Generic Models: State-of-the-art group recommendation models like SoAGREE performed poorly because they aggregate individual preferences without accounting for the specific structural and social context of OSS collaboration.
  • Architecture Validation: Ablation studies confirmed that the hierarchical three-layer structure is essential; "flatter" models (ALT1 and ALT2) that combined layers saw a marked decrease in recommendation quality.

Conclusion and Future Work

GAT-Team demonstrates that modeling social cohesion is critical for effective team recommendation in open-source environments. By treating teams as established social units rather than collections of individuals, the model provides high-quality, project-specific suggestions. Future research will address the "cold start" problem for new teams and explore the dynamic evolution of these social structures over time.

Summary updated

BibTeX
@inproceedings{qi2023gat,
  author = {Qing Qi and Jian Cao},
  title = {{GAT-Team: A Team Recommendation Model for Open-Source Software Projects}},
  booktitle = {SEKE},
  year = {2023},
  url = {https://openreview.net/forum?id=nAekT9iPRV},
}
Figure