CoRCi: Cross-Reconstruction of Coherent Interests Modeling in Cross-Domain Sequential Recommendation

Cross-Domain Sequential Recommendation (CDSR) aims to improve recommendation accuracy by transferring user interest patterns between domains. A persistent challenge is that existing methods often fragment user interests by training separate encoders for each domain or by merging sequences in a way that magnifies inter-domain discrepancies. The authors investigate how to better model domain-invariant interests while maintaining the coherence of user behavior across different domains.

The authors propose CoRCi (Cross-Reconstruction for Coherent Interest), a dual-target CDSR framework. Instead of training encoders on mixed-domain sequences from scratch, CoRCi uses a Cross-Reconstruction (CR) approach. This module generates mixed-domain representations directly from pre-encoded specific-domain sequences using cross-attention. To ensure these representations remain coherent, the authors introduce FocalNCE, a loss function that embeds Focal Loss into the standard InfoNCE objective. FocalNCE assigns higher penalties to negative samples drawn from the same domain as the query, effectively suppressing intra-domain bias and strengthening the alignment of domain-invariant interests.

Experimental results on four real-world datasets demonstrate that CoRCi consistently outperforms state-of-the-art CDSR methods across all evaluated metrics. The improvements are statistically significant, suggesting that the combination of cross-reconstruction and the FocalNCE loss effectively mitigates the fragmentation of user interests that plagues traditional dual-target training strategies. The model successfully captures evolving latent interests that span multiple domains, leading to more accurate next-item recommendations.

By addressing the fundamental tension between domain-specific and domain-invariant interest modeling, CoRCi provides a more robust way to leverage data from auxiliary domains. This is particularly valuable in real-world scenarios where data sparsity in a target domain limits the performance of single-domain models. The framework offers a scalable and effective alternative to existing CDSR techniques, providing a clear path for improving recommendation systems in multi-domain environments.