G³SR: Global Graph Guided Session-Based Recommendation.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Session-based recommendation systems utilize anonymous session data when user profiles are unavailable.
• Existing methods focus on intra-session item interactions, neglecting valuable historical relational information.
• This limitation hinders the ability to capture comprehensive user interests and improve recommendation quality.

Purpose of the Study:

• To propose a novel method, Global Graph Guided Session-based Recommendation (G3SR), to address the limitations of existing session-based recommendation approaches.
• To leverage historical relational information across all sessions for enhanced recommendation accuracy.
• To improve recommendations, particularly for cold items, by incorporating global item representations.

Main Methods:

• G3SR employs a two-step approach: building a global graph from all session data and learning unsupervised item representations.
• Learned global item representations are refined on session graphs using graph networks.
• A readout function generates session representations for each individual session.

Main Results:

• Extensive experiments on two real-world benchmark datasets demonstrate significant and consistent improvements of G3SR over state-of-the-art methods.
• The G3SR method shows particular effectiveness in improving recommendations for cold items.
• The approach successfully integrates global item information with session-specific context.

Conclusions:

• G3SR effectively enhances session-based recommendation by incorporating global item relational information.
• The proposed method offers a promising solution for improving recommendations, especially in scenarios with sparse or unavailable user profiles.
• G3SR demonstrates superior performance compared to existing methods, highlighting the value of global graph guidance.