Multi-channel high-order network representation learning research.

Area of Science:

• Graph Theory
• Machine Learning
• Network Science

Background:

• Existing network representation learning methods often rely on structural or external features.
• Capturing global network features is crucial for enhancing embedding quality and retaining comprehensive information.
• There's a need for algorithms that can effectively model high-order network structures.

Purpose of the Study:

• To propose a novel algorithm for multi-channel high-order network representation learning.
• To enhance node classification performance by preserving global structural features.
• To develop a method that efficiently utilizes network structure features for comprehensive modeling.

Main Methods:

• Developed the Multi-Channel High-Order Network Representation (MHNR) algorithm.
• Constructed high-order network features from the original network structure.
• Introduced a graph assimilation mechanism for single-channel learning within the multi-channel framework.
• Integrated multi-channel and single-channel mechanisms for joint high-order structure modeling.

Main Results:

• The MHNR algorithm demonstrated strong node classification performance on Citeseer, Cora, and DBLP datasets.
• MHNR outperformed existing comparison algorithms in node classification tasks.
• Optimized vector length resulted in up to 12.24% higher average classification accuracy compared to DeepWalk.

Conclusions:

• The proposed MHNR algorithm achieves optimal node classification performance using only network structural features.
• MHNR effectively models high-order network structures, improving embedding quality.
• The multi-channel approach enhances the utilization of network structural information.