A simple and effective convolutional operator for node classification without features by graph convolutional...

Area of Science:

• Graph Neural Networks
• Machine Learning
• Data Mining

Background:

• Graph Neural Networks (GNNs) excel in graph analysis by utilizing node features.
• Existing GNNs, including Graph Convolutional Networks (GCNs), struggle with graphs lacking node features.
• Node classification is a critical task in graph analysis.

Purpose of the Study:

• To develop a GNN approach for node classification on graphs without node features.
• To introduce a novel convolutional operator that enhances GNN performance.
• To explore the theoretical underpinnings of GNNs by examining their relationship with community detection.

Main Methods:

• Introduction of path-driven neighborhoods and an extensional adjacency matrix as a convolutional operator.
• Development of exopGCN, integrating the new operator into GCN for feature-less graphs.
• Application of the convolutional operator to 13 existing GNNs to assess its generalizability.

Main Results:

• exopGCN demonstrated superior performance in node classification on six real-world graphs lacking node features compared to other GNNs.
• Integrating the proposed convolutional operator significantly improved the accuracy of 13 different GNNs.
• A positive correlation was identified between node classification by GCNs without node features and community detection.

Conclusions:

• The proposed exopGCN effectively addresses node classification challenges in feature-less graphs.
• The novel convolutional operator offers a generalizable method for enhancing GNN accuracy across various architectures.
• The discovered link between node classification and community detection opens new theoretical avenues for GNN research.