Time Interval-Enhanced Graph Neural Network for Shared-Account Cross-Domain Sequential Recommendation.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Shared-account cross-domain sequential recommendation (SCSR) is crucial due to increasing multi-platform user accounts.
• Existing recurrent neural network (RNN)-based models struggle with multi-entity relationships and explicit cross-domain structures.

Purpose of the Study:

• To address limitations in current SCSR methods, particularly the lack of explicit cross-domain graph structures and time interval considerations.
• To develop a novel graph-based recommendation system that enhances representation learning for users and items across domains.

Main Methods:

• Proposed a time interval-enhanced domain-aware graph convolutional network (TiDA-GCN).
• Constructed domain-specific graphs linking users and items.
• Employed domain-aware graph convolutions with attention mechanisms for user-specific representations.
• Integrated time interval information and an account-aware self-attention module for enhanced learning.

Main Results:

• The TiDA-GCN model demonstrated superior performance in shared-account cross-domain sequential recommendation tasks.
• Effectively captured complex relationships among multiple entities and domains.
• Showcased the importance of incorporating time interval data and explicit graph structures.

Conclusions:

• The proposed TiDA-GCN offers a significant advancement in SCSR by leveraging graph structures and temporal information.
• This approach provides more accurate and discriminative representations for users and items.
• Highlights the potential of graph-based methods and temporal dynamics in recommendation systems.