Graph Spring Network and Informative Anchor Selection for session-based recommendation.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Mining

Background:

• Session-based recommendation (SBR) systems predict user interests in anonymous sessions.
• Existing methods struggle with capturing complex item relationships and learning effective ID-based item embeddings.
• Graph Neural Networks (GNNs) show promise but are not optimized for SBR's specific embedding learning needs.

Purpose of the Study:

• To develop a novel Graph Neural Network (GNN) for improved ID-based item embedding learning in SBR.
• To enhance the encoding of item relationships, especially for items distant in the session graph.
• To propose an integrated model for superior next-item prediction in anonymous sessions.

Main Methods:

• Introduced the Graph Spring Network (GSN) to optimize neighborhood affinity in item embeddings.
• Developed an Informative Anchor Selection (IAS) strategy using item entropy to identify key items.
• Integrated GSN and IAS into the GSN-IAS model, employing a gated recurrent unit (GRU) and adaptive fusion for prediction.

Main Results:

• The GSN-IAS model demonstrated superior performance compared to state-of-the-art methods on three public datasets.
• GSN effectively captures neighborhood affinity, improving ID-based item embedding quality.
• IAS successfully identifies informative anchors, enhancing the encoding of long-range item dependencies.

Conclusions:

• The proposed GSN-IAS model offers a significant advancement in session-based recommendation.
• The novel GSN architecture and IAS strategy effectively address limitations in current SBR approaches.
• GSN-IAS provides more accurate next-item predictions by better modeling item relationships and embeddings.