Abstract
Expert finding in Community Question Answering (CQA) platforms is essential for ensuring high-quality and timely responses. Traditional methods often rely on text-matching, which fails to capture complex collaborative signals and struggles with data sparsity among "long-tail" users. We present CGEF (Graph Collaborative Expert Finding), a framework that models expert-question interactions as a bipartite graph. By leveraging Graph Attention Networks (GAT) to capture high-order connectivity and employing a dual-view contrastive learning strategy (behavior-level and interest-level), CGEF learns robust representations even from sparse data. Evaluations on six StackExchange datasets demonstrate that CGEF significantly outperforms state-of-the-art baselines, particularly in precision and robustness for inactive experts.
Introduction
In Community Question Answering (CQA) platforms like StackOverflow or Zhihu, routing a new question to the right expert is crucial for high-quality, timely answers. Most existing models treat this as a text-matching problem, asking whether the text of a new question matches the text of questions a user has answered previously. However, this approach suffers from two major gaps: it ignores high-order connectivity (collaborative signals between users who share patterns but not vocabulary) and it fails to handle data sparsity for the majority of users who have answered very few questions.
The CGEF Framework
To address these challenges, we propose CGEF, which models the community as a bipartite graph of experts and questions. The framework consists of three primary components:
- Graph-Based Encoding: Using a Graph Attention Network (GAT), the model recursively propagates information across the graph. This allows it to learn latent connections—for example, if two questions are answered by the same expert, they are likely related.
- BERT Question Encoder: To handle "cold start" scenarios for new questions, a pre-trained BERT model encodes question titles into initial embeddings.
- Dual-View Contrastive Learning: To combat data sparsity, the model employs two augmentation tasks. Behavior-level augmentation randomly drops edges to force the model to learn essential interaction patterns. Interest-level augmentation replaces specific question titles with general tags (e.g., "Python") to capture broad expertise.
Key Insights: High-Order Connectivity
A core innovation of CGEF is the discovery that an expert is defined by the "company they keep." By looking multiple steps away in the graph (e.g., Question → Expert → Question → Expert), the model discovers expertise that text alone would miss. The GAT mechanism allows the model to assign higher weights to more informative neighbors when building an expert's profile.
However, there is a limit to this depth; the authors noted that going beyond three layers ($L=3$) leads to "over-smoothing," where nodes become indistinguishable as they aggregate too much information from distant parts of the graph.
Experimental Results
CGEF was evaluated on six real-world datasets from StackExchange, including AI, Bioinformatics, and History. The model consistently outperformed state-of-the-art baselines across all metrics, including MRR, P@1, and NDCG@10. Notably, on the "History" dataset, CGEF improved P@1 by over 22.9% compared to the best deep neural baseline. The model also proved exceptionally robust for "inactive" experts, where high-order graph connections successfully compensated for a lack of direct interaction data.
Conclusion
CGEF demonstrates that graph-based collaborative signals and contrastive learning are powerful tools for expert finding in sparse CQA environments. While the model currently treats the community as a static snapshot, future work may explore temporal dynamics to track how an expert's interests evolve over time.
Summary updated