Abstract
Community Question Answering (CQA) platforms have become vital knowledge-sharing hubs, yet they face significant challenges in efficiently matching questions with the right experts. This survey provides a comprehensive overview of expert recommendation techniques, categorizing methods from simple heuristics to complex hybrid models. We examine the current landscape of datasets, identify key insights regarding user expertise versus willingness, and highlight critical future directions such as dynamic modeling and solving the cold-start problem.
Introduction: The Matching Problem in CQA
CQA platforms like Stack Overflow and Yahoo! Answers rely on the "wisdom of the crowd." However, as these platforms scale to millions of users, a massive matching problem emerges: requesters often face long wait times or low-quality responses, while experts are overwhelmed by irrelevant queries. Expert recommendation systems, also known as question routing, aim to solve this by automatically identifying and notifying highly capable users for newly posted questions.
The CQA Ecosystem
To understand recommendation, one must first understand the underlying structure of CQA interactions. These platforms involve a complex web of relationships between requesters, answerers, and the content they produce.
As shown in the conceptual model above, users can take on roles as Requesters or Answerers (including specialized Experts). The system must track relations between Questions, Answers, and Feedback to determine who is best suited to provide a "Best Answer."
Categorization of Recommendation Methods
Research in this field has evolved through several distinct methodological phases:
- Simple Heuristics: Early methods used basic metrics like vote counts or TF-IDF similarity to match questions to a user's history.
- Language & Topic Models: Generative models like LDA and PLSA map questions to latent topics that users have historically engaged with.
- Network-based Methods: Algorithms like PageRank (e.g., ExpertiseRank) treat the platform as a graph, identifying authoritative "hubs" based on interaction edges.
- Supervised Learning: Classification and Learning-to-Rank approaches use SVMs or GBDTs to predict expert suitability based on extracted features.
- Collaborative Filtering: Matrix factorization techniques predict user-question affinity based on patterns across the entire user base.
- Hybrid Approaches: Modern state-of-the-art systems combine multiple strategies, such as merging topical relevance with network authority.
Datasets and Communities
The diversity of CQA platforms provides a rich set of environments for testing these algorithms. While Stack Overflow and Yahoo! Answers are the most common benchmarks, researchers utilize a wide array of specialized communities.
These communities vary significantly in language, domain (e.g., Medical, Tax, Programming), and quality control mechanisms, as seen in the table above. This variety is crucial for ensuring that recommendation models are robust across different contexts.
Key Insights and Challenges
Successful recommendation hinges on two pillars: Answer Probability (relevance and likelihood to respond) and User Expertise (quality of the potential answer). A common pitfall is confusing authoritativeness with willingness; the most knowledgeable experts are often the least available.
Furthermore, the "Cold Start" problem remains a significant hurdle. Most models rely on historical data, favoring established users and ignoring "rising stars." Additionally, most research treats user profiles as static, failing to account for interest drift or evolving expertise.
Future Directions
To advance the field, the survey suggests that future systems should:
- Incorporate Real-time Dynamics: Model user availability and psychological willingness to answer.
- Enable Collaborative Recommendations: Suggest groups of users with complementary skills for multi-faceted questions.
- Leverage External Data: Use signals from GitHub, blogs, or social media to mitigate the cold-start problem for new users.
- Adopt Modern Architectures: Integrate Deep Learning, Graph Embeddings, and Factorization Machines to better handle heterogeneous CQA data.
Summary updated