Hybrid Low-Order and Higher-Order Graph Convolutional Networks.

Area of Science:

• Graph representation learning
• Deep learning on graphs
• Network analysis

Background:

• Higher-order neighborhood information enhances graph representation learning accuracy.
• Existing higher-order graph convolutional networks suffer from high parameter counts and computational complexity.

Purpose of the Study:

• To propose a hybrid lower-order and higher-order graph convolutional network (HLHG) model.
• To reduce the number of network parameters and computational complexity in graph convolutional networks.

Main Methods:

• Implemented a weight sharing mechanism to decrease network parameters.
• Introduced a novel information fusion pooling layer to integrate high-order and low-order neighborhood matrix information.
• Conducted theoretical comparisons of computational complexity and parameter count against state-of-the-art models.

Main Results:

• The proposed HLHG model demonstrates reduced parameter count and computational complexity.
• Experimental validation on large-scale text and citation network datasets confirmed the model's effectiveness.
• Achieved higher classification accuracy with a minimal set of trainable weight parameters.

Conclusions:

• The HLHG model offers an efficient approach to graph representation learning.
• The hybrid strategy effectively balances the benefits of lower-order and higher-order information.
• This model presents a promising solution for accurate and computationally feasible graph classification tasks.