Triple contrastive learning framework for domain-level zero-shot recommendation

Domain-level zero-shot recommendation (DZSR) poses a significant challenge in recommender systems (RS), representing a cold-start scenario where the domain completely lacks interaction history. Previous works leveraged universal knowledge from item content to bridge domains, transferring user preferences from the source domain to the target domain. However, previous works relied solely on item content similarity for DZSR, while overlooking both co-interaction relevance and topic relevance. These limitations not only restrict the exploration of user interests, but also further hinder the transfer of user preferences across domains. To address limitations, this paper proposes a Triple Contrastive Learning Framework (TCLF) to perform three tailored alignments. Specifically, we first design a self-contrastive learning to align the embeddings generated from item descriptions and knowledge graph triples of the same item. Then, we measure the co-interaction relevance from both the user and item perspectives, and align co-interacted source items by a source-domain contrastive learning. Finally, we measure the topic relevance in an indirect manner, and design a cross-domain contrastive learning to align topic-relevant source and target items. Extensive experiments on four public datasets demonstrate the superiority of TCLF for DZSR tasks compared to state-of-the-art baselines.