Cross-domain sequential recommendation via interest-guided knowledge migration.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Recommender Systems

Background:

• Cross-domain sequence recommendation (CDSR) leverages auxiliary domains to address data sparsity.
• Existing CDSR methods struggle with integrating domain-invariant knowledge and capturing diverse user interests, leading to negative transfer.

Purpose of the Study:

• To propose a novel cross-domain sequential recommendation framework, Interest-Guided Knowledge Migration (IGKM).
• To effectively integrate domain-invariant knowledge and address the diversity of user interests in cross-domain scenarios.

Main Methods:

• Introduced the L²oME mechanism combining Low-Rank Adaptation (LoRA) and Mixture-of-Experts (MoE) for dynamic user representation refinement.
• Designed a domain-shared multi-interest extraction module to capture fine-grained user interests.
• Utilized interest representations as conditions for a cross-domain bridge transformer to enable selective knowledge migration.

Main Results:

• The IGKM framework demonstrated superior performance compared to state-of-the-art CDSR methods across three datasets.
• Effectively suppressed negative transfer, a common issue in cross-domain recommendation.
• Significantly improved recommendation accuracy.

Conclusions:

• The proposed IGKM framework offers an effective solution for cross-domain sequential recommendation.
• Interest-guided knowledge migration is crucial for mitigating negative transfer and enhancing recommendation performance.
• The framework's ability to capture and utilize domain-specific user interests leads to more accurate personalized recommendations.