A Knowledge-Graph-Enhanced Conversational Recommender System Based on Graph Augmentation

Conversational Recommender Systems (CRSs) have emerged as a new paradigm for personalized recommendations by engaging users in dialogue to refine their preferences dynamically. This interaction-driven approach is particularly effective in cold-start scenarios, where user interests are initially uncertain. Generally, CRSs integrate a language module for understanding user utterances and a knowledge graph (KG) module for providing structured entity information. While KGs are important for improving recommendation quality, their inherent sparsity limits the effectiveness of existing graph neural network (GNN)based methods, leading to suboptimal performance. To address this challenge, we propose GAKGE-CRS, a Knowledge-GraphEnhanced Conversational Recommender System with Graph Augmentation. Our method enhances KG representation by introducing virtual nodes to mitigate sparsity and improve recommendation accuracy. Firstly, we propose a graph augmentation approach to generate virtual neighbor nodes, enriching the structural connectivity of KGs. Secondly, we enhance R-GCN's neighbor aggregation by adaptively integrating real and virtual neighbors to refine node representations. Thirdly, a pretraining strategy is used to optimize the proposed model, ensuring effective virtual neighbor generation during fine-tuning. Experiments on widely used CRS datasets demonstrate that our approach outperforms existing methods, highlighting the effectiveness of graph augmentation in CRSs.