LGNN: a novel linear graph neural network algorithm.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Graph Neural Networks

Background:

• Deep learning excels in image recognition and graph neural networks (GNNs).
• Existing GNNs capture local graph structures via spatial or spectral domains, requiring significant computation.
• Modeling high-order network characteristics often necessitates complex deep or multi-channel network structures.

Purpose of the Study:

• Propose a novel Linear Graph Neural Network (LGNN) framework.
• Enhance computational efficiency and model high-order graph structures.
• Evaluate LGNN's performance against existing GNN algorithms.

Main Methods:

• Input graph preprocessing using symmetric and feature normalization.
• High-order adjacency matrix propagation for iterative neighbor feature aggregation.
• Simple linear mapping for efficient final node representation generation.

Main Results:

• LGNN achieves performance comparable to or exceeding mainstream GNNs in most evaluation tasks.
• LGNN shows particular strength on sparse network datasets.
• LGNN's performance is slightly lower than some existing algorithms on specific tasks.

Conclusions:

• LGNN offers a computationally efficient approach to modeling high-order graph structures.
• The proposed framework provides a viable and effective alternative for various graph neural network applications.
• LGNN demonstrates strong performance, particularly in scenarios with sparse graph data.