A Dual-Expert Strategy Integrating LLMs to Mitigate Negative Transfer in Cross-Domain Sequential Recommendation

How can Large Language Models (LLMs) be adapted for Cross-Domain Sequential Recommendation (CDSR) to leverage rich textual features while avoiding the performance degradation caused by negative transfer across disparate domains?

The authors propose DuELRec, an ID-free LLM-based recommender system designed to mitigate negative transfer. The model first processes user interaction sequences using an LLM backbone with instruction tuning. It then introduces a domain-gated dual-expert framework comprising two parallel transformer-based experts: a single-domain expert that restricts attention to items within the same domain, and a cross-domain expert that allows attention across all domains. An item-aware attention transformation module aggregates textual subtokens into item-level representations and applies block-level causal masks. A gating network adaptively combines the outputs of these two experts, using single-domain signals to suppress cross-domain noise. Finally, a dual-sampling token-to-item contrastive learning objective is applied using stochastic negative sampling from both single- and cross-domain item pools to better capture item-level collaborative signals.

Evaluated on two real-world datasets spanning ten domains, DuELRec outperforms 26 state-of-the-art ID-based and LLM-based recommendation baselines. Furthermore, deploying the model in a commercial personal assistant app resulted in a 47.6% relative increase in the click-through rate, demonstrating significant real-world efficacy.

Cross-domain recommendation systems often suffer when knowledge transferred between domains interferes rather than helps. By resolving this negative transfer using dual-expert gating and text-based contrastive learning, this work demonstrates that pure LLM-based recommenders can surpass traditional ID-based models without requiring domain-specific retraining tables.