Mitigating Matthew Effect: Multi-Hypergraph Boosted Multi-Interest Self-Supervised Learning for Conversational Recommendation
Abstract Overview
This paper studies the Matthew effect in conversational recommender systems, where repeated user-system interactions can increasingly favor popular items and narrow users' exposed interests over time. The authors propose HiCore, an end-to-end framework that learns multi-level user interests through multiple hypergraphs built from three channels: group, joint, and purchase behaviors. HiCore constructs item-oriented, entity-oriented, and word-oriented hypergraphs, then applies self-supervised learning to aggregate these signals into interest representations used by both recommendation and conversation modules. The empirical study covers four conversational recommendation benchmarks and evaluates recommendation quality, conversational diversity, and metrics related to mitigating popularity concentration.
Novelty
The paper's main novelty is framing Matthew-effect mitigation directly within conversational recommendation, rather than only in static recommendation settings. It introduces a multi-hypergraph, triple-channel design that jointly models item-, entity-, and word-level interests, and couples this with self-supervised learning so the learned interests can support both recommendation and response generation.
Results
Across REDIAL, TG-REDIAL, OpenDialKG, and DuRecDial, the reported results show HiCore outperforming compared baselines on recommendation metrics and conversational diversity metrics, with several tables marking statistically significant gains. In the Matthew-effect analysis, HiCore achieves the highest coverage and the lowest reported average popularity and long-tail ratio values among the compared methods, indicating reduced concentration on popular items. Ablation experiments further show that removing any hypergraph component degrades recommendation performance.
Key Points
- HiCore models user interests with multiple hypergraphs spanning item, entity, and word views under group, joint, and purchase channels.
- The learned multi-interest representation is used in both the recommendation module and the conversation module, linking interest modeling to dynamic conversational interactions.
- Experiments and ablations indicate stronger recommendation and conversational performance than baselines, while diversity-oriented analyses suggest improved mitigation of the Matthew effect.