An Ensemble Learning Approach to Graph Learning Based on Evolutionary Graph Neural Architecture Search

Area of Science:

• Artificial Intelligence
• Machine Learning
• Graph Neural Networks

Background:

• Graph Neural Networks (GNNs) excel in non-Euclidean data mining.
• Neural Architecture Search (NAS) is used to optimize GNNs, but existing methods struggle with high-dimensional data and search space exploration.
• Current graph NAS (GNAS) can lead to local optima and underutilized search potential.

Purpose of the Study:

• To address limitations in existing GNAS algorithms, particularly in handling high-dimensional data and optimizing search space utilization.
• To propose a novel framework, Ensemble-GNAS, integrating evolutionary GNAS with ensemble learning.
• To enhance the applicability and accuracy of GNNs through automated architecture search.

Main Methods:

• Developed an evolutionary GNAS method incorporating a niche strategy to improve local search and diversify candidate networks.
• Implemented an ensemble fusion strategy using a tree-structured Parzen estimator (TPE) to optimize weights of candidate networks.
• Evaluated the framework on node classification, link prediction, and graph classification datasets.

Main Results:

• Ensemble-GNAS demonstrated superior performance compared to state-of-the-art methods across various graph mining tasks.
• The approach achieved higher classification accuracy and improved performance stability.
• Ablation studies confirmed the effectiveness of the integrated niche and TPE strategies.

Conclusions:

• Ensemble-GNAS effectively overcomes limitations of traditional GNAS by enhancing search capability and ensemble optimization.
• The proposed framework offers a robust solution for optimizing GNN architectures in complex graph data mining.
• The method shows significant promise for improving GNN performance in diverse applications like biological network analysis and citation network classification.