On Mitigating Data Sparsity in Conversational Recommender Systems

Abstract

Conversational recommender systems (CRSs) infer user preferences from dialogue contexts, but they suffer from severe data sparsity in both dialogue and entity spaces. Dialogue data are linguistically diverse and open-ended, making it difficult to generalize across varied expressions. Meanwhile, existing CRS models often rely on large knowledge graphs, where only a small fraction of entities receive effective supervision during training, leaving the majority under-trained or entirely unseen at inference time. To address these challenges, we propose DACRS, a novel CRS framework consisting of three modules: Dialogue Augmentation, Knowledge-Guided Entity Modeling, and Dialogue-Entity Matching. The Dialogue Augmentation module adopts a two-stage augmentation pipeline to enrich dialogue contexts and improve robustness to linguistic variation. The Knowledge-Guided Entity Modeling module leverages knowledge graphs through entity substitution and an entity similarity constraint to enhance representation learning for sparsely supervised and unseen entities. Finally, the Dialogue-Entity Matching module integrates dialogue representations with mentioned entity embeddings via dialogue-guided attention aggregation, yielding user representations that capture both explicit and implicit preferences. Extensive experiments on two public benchmark datasets demonstrate that DACRS consistently outperforms state-of-the-art conversational recommender systems.