Convolution Bridge: An Effective Algorithmic Migration Strategy From CNNs to GNNs.

Area of Science:

• Machine Learning
• Graph Neural Networks
• Computer Vision

Background:

• Graph neural networks (GNNs) excel in relational data but face challenges in cross-domain model migration.
• Migrating models from domains like computer vision (CV) to GNNs often requires extensive reconstruction.
• Preserving convolutional properties during migration is crucial for optimizing GNN development.

Purpose of the Study:

• To propose a novel 'convolution bridge' for efficient migration of Convolutional Neural Network (CNN) models to GNNs.
• To facilitate data alignment between CNN and GNN architectures.
• To enable effective transfer of CNN-based models to graph-structured data.

Main Methods:

• Developed a 'convolution bridge' to align data structures between CNNs and GNNs.
• Migrated Inception module to GraInc for node-level tasks.
• Migrated U-Net architecture to GraU-Net for graph-level tasks.

Main Results:

• The proposed convolution bridge enables efficient CNN to GNN model migration.
• Migrated models GraInc and GraU-Net demonstrate competitive performance against state-of-the-art GNNs.
• Performance gains are particularly notable on dense graph datasets.

Conclusions:

• The convolution bridge is an effective strategy for migrating CNN architectures to GNNs.
• GraInc and GraU-Net show promise for node-level and graph-level tasks, respectively.
• This approach simplifies cross-domain model transfer in machine learning.