Revisiting 2D Convolutional Neural Networks for Graph-Based Applications.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Graph convolutional networks (GCNs) face implementation challenges with irregular graph inputs.
• Convolutional neural networks (CNNs) excel at feature extraction but lack support for general graph data.

Purpose of the Study:

• To develop methods for mapping general graphs to 2D grids compatible with CNNs.
• To preserve graph topology during the graph-to-grid mapping process for effective analysis.

Main Methods:

• Proposed two novel graph-to-grid mapping schemes: Graph-Preserving Grid Layout (GPGL) and Hierarchical GPGL (H-GPGL).
• Formulated GPGL as integer programming and developed an approximate solver using a penalized Kamada-Kawai method with a vertex separation penalty.
• Integrated the graph-to-grid mapping with 2D CNN architectures like VGG16, ResNet50, and Multi-Scale Maxout (MSM) CNN.

Main Results:

• Demonstrated the effectiveness of GPGL for general graph classification on small graphs.
• Showcased the performance of H-GPGL for 3D point cloud segmentation on large graphs using 2D CNNs.
• The proposed methods enable CNNs to leverage graph structural information.

Conclusions:

• GPGL and H-GPGL successfully bridge the gap between GCNs and CNNs by enabling graph data processing with CNNs.
• These novel mapping schemes offer efficient and effective solutions for graph-based machine learning tasks.
• The approach enhances capabilities in graph classification and 3D point cloud segmentation.