In this paper, we present a novel approach to the multi-target cross-domain recommendation (MTCDR) problem, Generative Multi-target Cross-domain Recommendation (GMC). Unlike existing MTCDR methods that rely on shared entities across domains, GMC uses an item tokenizer to generate domain-shared semantic IDs, which are then used to integrate multi-domain knowledge into a domain-integrated generative model. We formalize item recommendation as a tokenization task, train a domain-integrated sequence-to-sequence model, and improve its performance with domain-aware contrastive loss and domain-specific fine-tuning. Experimental results on five public datasets demonstrate that GMC outperforms existing methods.
