Locality preserving dense graph convolutional networks with graph context-aware node representations.

Area of Science:

• Graph Neural Networks
• Machine Learning
• Data Science

Background:

• Graph Convolutional Networks (GCNs) are effective for graph representation learning.
• Existing GCNs struggle to preserve crucial local graph information, hindering graph classification performance.

Purpose of the Study:

• To propose a novel locality-preserving dense GCN model.
• To enhance graph classification by improving local information preservation and node representativeness.

Main Methods:

• Incorporated a local node feature reconstruction module using an encoder-decoder mechanism.
• Introduced dense connectivity between convolutional layers and readouts to capture multi-range local structures.
• Utilized a self-attention module to aggregate layer-wise representations for a final graph-level representation.

Main Results:

• The proposed model demonstrated superior performance in graph classification tasks.
• Achieved higher classification accuracy compared to existing state-of-the-art methods on benchmark datasets.

Conclusions:

• The locality-preserving dense GCN effectively addresses the limitations of existing GCNs.
• The model's ability to preserve local information and generate context-aware representations leads to improved graph classification accuracy.