GDCNet: Graph Enrichment Learning via Graph Dropping Convolutional Networks.

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Graph convolutional networks (GCNs) are pivotal for graph data representation and learning.
• Traditional GCNs often use a single, fixed spatial convolution filter, limiting their capacity to capture complex graph patterns.
• This contrasts with Convolutional Neural Networks (CNNs) that utilize diverse filters for image data.

Purpose of the Study:

• To enhance the feature extraction capabilities of GCNs for complex graph data.
• To introduce a novel network architecture that overcomes the limitations of fixed filters in GCNs.
• To improve the representation learning capacity for graph-structured data.

Main Methods:

• Proposed a graph-dropping convolution layer (GDCLayer) inspired by depthwise separable convolution and DropEdge.
• GDCLayer generates various graph convolution filters by randomly dropping edges from the input graph.
• Developed a new end-to-end network architecture, the graph-dropping convolutional network (GDCNet), utilizing GDCLayer.

Main Results:

• The proposed GDCNet demonstrates effectiveness in graph data learning tasks.
• Experiments on multiple datasets validate the superior performance of the GDCNet architecture.
• The use of GDCLayer leads to richer feature descriptors for graph data.

Conclusions:

• The novel GDCNet architecture effectively addresses the limitations of fixed filters in traditional GCNs.
• The proposed graph-dropping convolution approach enhances the ability to encode complex patterns in graph data.
• GDCNet offers a promising advancement for representation learning on graph-structured data.