Graph Kernel Neural Networks.

Area of Science:

• Machine Learning
• Graph Neural Networks
• Computer Science

Background:

• Convolution operators are fundamental to neural networks, excelling with grid-like data (e.g., images).
• Extending convolutions to irregular graph structures presents significant challenges.
• Existing methods often rely on graph embeddings, which can be complex and computationally intensive.

Purpose of the Study:

• To develop a novel approach for extending the convolution operator to graph domains.
• To define a purely structural neural network model for graph data.
• To enhance interpretability of graph neural network models.

Main Methods:

• Utilizing graph kernels to define a convolution operator on graphs.
• Developing an architecture that integrates pluggable graph kernels.
• Conducting an extensive ablation study to analyze hyperparameter impact.

Main Results:

• The proposed method successfully extends convolution to graph structures.
• The model demonstrates interpretability through learned structural masks.
• Achieved competitive performance on standard graph classification and regression datasets.

Conclusions:

• Graph kernels provide an effective mechanism for structural convolutions on graphs.
• The proposed architecture offers a flexible and interpretable alternative for graph-based deep learning.
• The model shows promise for various graph analysis tasks.