HyCoRec: Hypergraph method reduces popularity bias in conversational recommendations

The Matthew effect (rich get richer, poor get poorer) is a persistent bias in recommender systems, where popular items become overexposed while niche items are ignored. Most existing methods only tackle this in static scenarios, but the problem worsens as users interact over time. To address this, a team of researchers from multiple Chinese institutions has introduced HyCoRec (Hypergraph-Enhanced Multi-Preference Learning for Alleviating Matthew Effect in Conversational Recommendation).

HyCoRec is specifically designed for conversational recommendation, where the system chats with users to refine suggestions. It learns from five preference aspects: item-, entity-, word-, review-, and knowledge-based. By modeling these diverse signals through a hypergraph structure, HyCoRec captures complex, higher-order relationships that standard graphs miss. This allows the model to surface less popular items without sacrificing recommendation relevance.

In extensive experiments on two benchmark datasets, HyCoRec achieves state-of-the-art performance in both recommendation accuracy and fairness metrics. It significantly reduces the gap between popular and niche item exposure while maintaining conversational quality. The code is available on GitHub, enabling further research and practical deployment.