Cross-domain sequential recommendation via interest-guided knowledge migration

Cross-domain sequence recommendation (CDSR) has gained widespread attention through leveraging information from auxiliary domains to alleviate data sparsity issues in the target domain. However, most existing CDSR methods fail to effectively integrate domain-invariant knowledge into domain-specific models. Additionally, they often overlook the diversity and domain-specificity of user interests in cross-domain scenarios, which may lead to negative transfer. To address these challenges, we propose a cross-domain sequential recommendation framework based on interest-guided knowledge migration (IGKM). First, we introduce the L²oME mechanism, which combines Low-Rank Adaptation (LoRA) with a Mixture-of-Experts (MoE) architecture and dynamically refines user domain-specific representations through layer-wise ensemble operations. Then, we design the domain-shared multi-interest extraction module to capture fine-grained user interest representations from cross-domain shared sequences. Next, we utilize these interest representations as interest-guided conditions and integrate them with the cross-domain bridge transformer module, enabling selective knowledge migration and personalized recommendation. Experimental results on three datasets demonstrate that the IGKM framework outperforms state-of-the-art cross-domain recommendation methods by effectively suppressing negative transfer and improving recommendation accuracy. The source code is publicly available at https://github.com/userLi2001/IGKM.git.