Empowering Cross-Domain Sequential Recommendation with Hybrid Tokenization and Serial-Parallel Decoding

The paper introduces GenCDSR, a generative framework for cross-domain sequential recommendation. Its hybrid tokenization uses a multi-tower architecture with hierarchical shared-specific and fine-grained codebooks to model both cross-domain commonalities and domain-specific distinctions. A serial-parallel decoding strategy exploits the hierarchical semantic identifiers to partially parallelize generation while maintaining consistency. On three public datasets, GenCDSR reportedly improves average accuracy by 1.5% and reduces average inference latency by 85.1% versus state-of-the-art baselines. Code and datasets are provided through the project repository.