Node-personalized multi-graph convolutional networks for recommendation.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Recommender Systems

Background:

• Graph neural networks (GNNs) show promise in recommendation systems.
• Existing GNNs often overlook the heterogeneous nature of user-item bipartite graphs.
• Differentiating node types is crucial for learning effective representations.

Purpose of the Study:

• To develop a novel node-personalized multi-graph convolutional network (NP-MGCN) for ranking recommendation.
• To address the limitations of homogeneous graph assumptions in existing methods.
• To enhance node representation by accounting for heterogeneity.

Main Methods:

• Proposed a node importance awareness block using node degree information.
• Developed a graph construction module fusing Jaccard similarity and co-occurrence matrices for user-user and item-item graphs.
• Designed a composite hop framework with single-hop (heterogeneous) and double-hop (homogeneous) branches for information propagation and aggregation.

Main Results:

• NP-MGCN generates more discriminative user and item node embeddings.
• The model effectively integrates heterogeneity of different nodes.
• Experimental results demonstrate superior recommendation performance compared to existing methods on multiple datasets.

Conclusions:

• NP-MGCN offers a significant advancement in ranking recommendation by embracing graph heterogeneity.
• The proposed architecture effectively captures complex user-item relationships.
• This approach provides a more robust and accurate recommendation framework.