ScaleGCN: Efficient and Effective Graph Convolution via Channel-Wise Scale Transformation.

Area of Science:

• Machine Learning
• Graph Neural Networks
• Deep Learning

Background:

• Graph convolutional networks (GCNs) effectively combine node features and graph topology for expressive embeddings.
• Traditional GCNs utilize neighborhood aggregation and fully-connected (FC) layers, but this coupling risks overfitting with larger receptive fields.
• FC layers can mix and pollute features across channels, hindering convergence and introducing noise.

Purpose of the Study:

• To investigate graph convolution methods that eliminate the need for FC layers.
• To propose and evaluate a novel graph convolution technique, scale graph convolution (SGC).
• To demonstrate SGC's advantages in reducing overfitting, improving convergence, and lowering computational/memory costs.

Main Methods:

• Developed scale graph convolution (SGC), a new graph convolution layer.
• SGC employs channel-wise scale transformation for node feature extraction, avoiding FC layers.
• Theoretical analysis and empirical validation were conducted to assess SGC's performance.

Main Results:

• SGC demonstrates a lower risk of overfitting compared to traditional GCNs.
• Fewer layers are required for models using SGC to converge.
• Models with SGC exhibit reduced computational and memory costs.

Conclusions:

• Scale graph convolution offers a cost-effective and efficient alternative to traditional GCN architectures.
• SGC achieves state-of-the-art performance across various datasets.
• Eliminating FC layers through channel-wise scaling is a promising direction for GCN research.