A multi-agent reinforcement learning framework for cross-domain sequential recommendation.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Recommender Systems

Background:

• Sequential recommendation models predict user behavior based on historical interactions.
• Data sparsity and misaligned user interests across domains challenge current cross-domain recommendation approaches.
• Existing methods often overlook collaborative transfer strategies, limiting performance.

Purpose of the Study:

• To propose a novel multi-agent reinforcement learning framework (MARL4CDSR) for cross-domain sequential recommendation.
• To address challenges of data sparsity and domain interest misalignment in sequential recommendations.
• To improve recommendation accuracy by optimizing knowledge transfer across domains.

Main Methods:

• Developed a multi-agent reinforcement learning framework (MARL4CDSR) where agents select relevant source domain items for transfer.
• Implemented an information fusion module with cross-attention to align item embeddings between source and target domains.
• Utilized a reward function based on next-item score differences to optimize the multi-agent system.

Main Results:

• MARL4CDSR significantly outperformed all baseline models across evaluated metrics on three Amazon domains.
• Demonstrated substantial improvements in NDCG@10 (14.76%) and HR@10 (10.25%) for the Movies&Books→Toys task, particularly in sparse target domains.
• The agent-based item selection and cross-attention fusion effectively enhanced recommendation performance.

Conclusions:

• MARL4CDSR offers a robust solution for cross-domain sequential recommendation, effectively handling data sparsity and domain misalignment.
• The multi-agent approach enables optimized and collaborative knowledge transfer, leading to superior recommendation quality.
• This framework represents a significant advancement in leveraging cross-domain user data for personalized recommendations.