HiCore framework tackles Matthew effect in conversational recommenders
Multi-hypergraph learning stops popular items from dominating recommendations
Conversational recommender systems (CRSs) face a persistent problem: the Matthew effect, where popular items get recommended more and even popular ones get ignored. Traditional mitigation methods work for static scenarios but fail in dynamic user-system interactions. To solve this, researchers from multiple institutions introduce HiCore, a framework that uses multi-hypergraph boosted multi-interest self-supervised learning. HiCore builds three hypergraphs—focused on items, entities, and words—to capture diverse user interests across different semantic channels, ensuring that niche but relevant items surface.
HiCore then leverages self-supervised learning to refine these interest representations, reducing the bias toward popular items while maintaining overall recommendation quality. Extensive tests on four CRS benchmark datasets show that HiCore achieves new state-of-the-art results, outperforming existing methods in both fairness and accuracy. The code is publicly available, enabling the research community to build on this work. This approach offers a practical way to make conversational AI assistants more equitable, ensuring users discover a wider range of relevant content.
- HiCore constructs three hypergraphs (item, entity, word) to model multi-level user interests and combat popularity bias
- Uses self-supervised learning to dynamically adapt recommendations during conversational feedback loops
- Achieves state-of-the-art results on four conversational recommendation datasets, balancing accuracy and fairness
Why It Matters
Makes conversational recommenders fairer by surfacing less popular but relevant content, improving long-term user engagement.