GenCDSR: Hybrid Tokenization for Cross-Domain Sequential Recommendation

A recent paper published on arXiv (2607.28659) presents GenCDSR, a novel generative framework designed for cross-domain sequential recommendation (CDSR). This approach tackles two significant challenges found in current generative recommendation methods: the oversight of collaborative relationships across domains during tokenization and the use of inefficient decoding techniques, such as beam search, which impede real-time application. GenCDSR utilizes a hybrid tokenization method that spans domains, featuring a multi-tower architecture to effectively capture both shared and unique characteristics through hierarchical shared-specific and detailed codebooks. Additionally, the authors introduce a serial-parallel decoding strategy to enhance efficiency. This work is classified as a 'new' announcement type on arXiv and can be accessed via the provided URL.