Mitigating Matthew Effect: Multi-Hypergraph Boosted Multi-Interest Self-Supervised Learning for Conversational Recommendation

Conversational Recommender Systems (CRSs) often suffer from the Matthew effect, where popular items receive disproportionate attention, leading to echo chambers and reduced discovery of niche content. While existing methods address this in static settings, they often fail to account for the dynamic feedback loops inherent in ongoing user-system conversations. This paper asks how to effectively model multi-level user interests to mitigate this bias in a conversational context.

The authors propose HiCore, a framework that integrates multi-hypergraph learning with a conversational recommendation engine. HiCore constructs triple-channel hypergraphs (group, joint, and purchase channels) across three levels: item-oriented (capturing social and behavioral patterns), entity-oriented (leveraging DBpedia for semantic relationships), and word-oriented (using ConceptNet to understand conversational intent). These hypergraphs are processed via a hypergraph convolution network to generate comprehensive, multi-interest user embeddings. These embeddings are then used to guide both the recommendation task (predicting items) and the conversation task (generating natural language responses) through a self-supervised learning objective.

By explicitly modeling user interests through diverse, high-order relationships rather than simple user-item interactions, HiCore provides a more nuanced understanding of user intent. This approach effectively breaks the cycle of popularity bias, allowing the system to suggest relevant but less-frequently-recommended items. The framework demonstrates that integrating multi-level interest modeling into the conversational loop is a viable strategy for improving both recommendation diversity and the quality of human-like dialogue.