Hypergraph Conversational Recommendation System Fusing Pairwise Relationships

Conversational recommendation systems aim to provide high-quality recommendations based on user needs through multiple rounds of interaction with users. Hypergraphs are introduced into conversation recommendation due to their ability to express and model complex relationships among multiple entities, enabling the capture of complex multi-entity interactions in dialog history. However, existing hypergraph-based methods treat all entities within the same hyperedge as sharing a single relationship, ignoring the fact that multiple types of semantic relationships coexist among entities within the same hyperedge. This leads to ambiguous entity representations and makes it difficult to accurately characterize complex user preferences. To address this issue, this paper proposes a Hypergraph Conversational Recommendation System Fusing Pairwise Relationships (HCRS-PR) model that integrates pairwise relationships. While preserving the overall high-order semantics of the hypergraph, it constructs a fine-grained pairwise relationship graph for each entity interaction within a hyperedge, capturing specific interaction patterns between entities and significantly improving the accuracy of conversational context representation. During the model inference stage, to enhance the diversity of generated responses, this paper adopts a multinomial beam search strategy based on multinomial distribution sampling. Experimental results on benchmark datasets demonstrate the effectiveness of the proposed method in conversation recommendation tasks.