PSA-GNN: An augmented GNN framework with priori subgraph knowledge.

Area of Science:

• Graph Neural Networks
• Machine Learning
• Network Science

Background:

• Graph Neural Networks (GNNs) are a dominant paradigm for graph representation learning, relying on iterative message passing among neighboring nodes.
• Current GNNs insufficiently leverage higher-order subgraph structures (e.g., motifs, cliques, cores) which hold crucial information for downstream tasks.
• Traditional graph research has extensively studied these subgraphs for tasks like node classification.

Purpose of the Study:

• To extend the receptive field and enhance the expressive power of GNNs beyond the limitations of 1st-order neighborhood aggregation.
• To integrate pre-mined subgraph structures as prior knowledge into GNN frameworks.
• To develop a GNN model that effectively utilizes both node-level and subgraph-level information.

Main Methods:

• Introduction of the Priori Subgraph Augmented Graph Neural Network (PSA-GNN) framework.
• Augmenting GNN layers with parallel convolution layers operating on a bipartite graph connecting nodes and pre-mined subgraphs.
• Incorporating trainable subgraph embeddings and weights, and introducing a homogeneity metric for subgraph purification during training.

Main Results:

• PSA-GNN establishes a hybrid receptive field by treating subgraphs as extended neighbors.
• The model demonstrates improved performance compared to state-of-the-art message passing GNNs.
• Effective purification of noisy subgraphs is achieved both heuristically and deterministically.

Conclusions:

• PSA-GNN significantly enhances GNNs' ability to capture complex graph structures by integrating prior subgraph information.
• The proposed framework offers a more powerful alternative to standard message passing GNNs, surpassing the 1st-order Weisfeiler-Lehman isomorphism test.
• PSA-GNN provides a flexible and effective method for leveraging structural priors in graph representation learning.