GenCDSR boosts cross-domain recommendations: 85% faster with hybrid tokenization

Cross-domain sequential recommendation (CDSR) tries to model how user interests shift across different platforms or categories. Existing generative recommendation (GR) methods learn semantic identifiers (SIDs) from item content and then treat the next-item prediction as a text-generation task. But two bottlenecks hurt them: they fail to capture collaborative correlations across domains during tokenization, and they rely on slow decoding strategies like beam search that aren't practical for real-time systems.

To solve that, researchers at City University of Hong Kong propose GenCDSR, a new framework that redesigns both tokenization and decoding. The cross-domain hybrid tokenization uses a multi-tower architecture with hierarchical shared-specific codebooks, so it learns both what's common across domains and what's unique to each. For decoding, the serial-parallel strategy exploits the hierarchical structure of SIDs to generate multiple parts of the sequence in parallel, cutting inference time dramatically while keeping output consistent. On three public benchmarks, GenCDSR achieves a 1.5% average accuracy gain and an 85.1% reduction in inference latency compared to state-of-the-art baselines. The authors note the Pareto improvement in both quality and speed, making it viable for real-time deployment. The paper was accepted to RecSys 2026, with code and datasets released online.