Adaptive multi-graph contrastive learning for bundle recommendation.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Bundle recommendation systems are gaining traction, utilizing Graph Neural Networks (GNNs) to model user-item interactions.
• Existing GNN models struggle to capture complex ternary relationships and suffer from noise due to disparate graph combinations.

Purpose of the Study:

• To propose a novel approach, Adaptive Multi-Graph Contrastive Learning for Bundle Recommendation (AMCBR), to address limitations in current bundle recommendation models.
• To enhance the modeling of intricate ternary relationships and improve the robustness of combined graph embeddings.

Main Methods:

• AMCBR constructs multiple graphs: a bundle preference graph, a collaborative neighborhoods graph, and an item-level preference hypergraph.
• It employs (hyper)graph convolution for embedding generation and an adaptive aggregation module for robust fusion of embeddings from different graphs.
• A contrastive learning strategy is utilized for joint model optimization and strengthening inter-graph collaborative links.

Main Results:

• AMCBR effectively models ternary interactions and mitigates noise through adaptive aggregation.
• The contrastive learning strategy strengthens collaborative links between individual graphs.
• Experiments show AMCBR outperforms state-of-the-art baselines in Top-K bundle recommendations on three real datasets.

Conclusions:

• AMCBR offers a robust and effective solution for bundle recommendation by capturing complex relationships and adaptively integrating information from multiple graph structures.
• The proposed adaptive multi-graph contrastive learning framework significantly advances the state-of-the-art in personalized bundle recommendations.