GenCDSR Boosts Cross-Domain Recommendation Accuracy and Speed

Cross-domain sequential recommendation (CDSR) systems aim to understand and predict user interests and behaviors across different product or content domains. While generative recommendation (GR) approaches have shown promise by converting item semantics into "semantic identifiers" (SIDs) and using autoregressive generation, current methods face two key limitations. They often neglect collaborative correlations between domains during the tokenization process, and their decoding strategies, such as beam search, are inefficient for real-time deployment. To overcome these challenges, researchers developed GenCDSR, an effective and efficient generative framework for CDSR. GenCDSR incorporates a novel cross-domain hybrid tokenization mechanism, featuring a multi-tower architecture. This design allows the system to simultaneously capture both shared commonalities and domain-specific distinctions through hierarchical shared-specific and fine-grained codebooks, ensuring a richer understanding of item semantics across domains. Furthermore, GenCDSR introduces a cross-domain serial-parallel decoding strategy. This innovative approach leverages the hierarchical structure of SIDs to partially parallelize the generation process. This significantly reduces inference latency while maintaining the consistency of the generated recommendations. Experiments on three public datasets demonstrated that GenCDSR achieved an average accuracy improvement of 1.5% and an impressive average inference latency reduction of 85.1% compared to state-of-the-art baselines, making it highly suitable for real-time applications.