Automatic Design of Deep Graph Neural Networks With Decoupled Mode

Area of Science:

• Artificial Intelligence
• Machine Learning
• Graph Neural Networks

Background:

• Graph Neural Networks (GNNs) are effective for node classification on graph data.
• Shallow GNNs have limitations, leading to research in deep GNNs.
• Manual design of deep GNNs is challenging due to issues like over-smoothing.

Purpose of the Study:

• To propose a novel neural architecture search (NAS) method for automated deep GNN design.
• To address limitations in current GNN architecture search for deep networks.
• To enhance node classification performance on large-scale graph data.

Main Methods:

• Redesigned the GNN search space using a decoupled propagation and transformation mode.
• Formulated the architecture search as a multi-objective optimization problem.
• Balanced accuracy and computational efficiency.

Main Results:

• The proposed NAS method successfully designs deep GNNs automatically.
• Achieved strong performance on various node classification tasks across benchmark datasets.
• Demonstrated scalability on large-scale graph datasets.

Conclusions:

• The novel NAS method effectively automates deep GNN design.
• The approach overcomes challenges associated with manual deep GNN design.
• The method is scalable and performs well on diverse node classification applications.