GeCNs: Graph Elastic Convolutional Networks for Data Representation.

Area of Science:

• Machine Learning
• Graph Representation Learning

Background:

• Graph Convolutional Networks (GCNs) are powerful for graph representation and learning.
• Standard GCNs aggregate features from all neighbors, which can be suboptimal and sensitive to noisy graph structures.

Purpose of the Study:

• To propose a novel Graph Elastic Convolution (GeC) operation for improved GCN learning.
• To enable adaptive neighbor selection during feature aggregation in GCNs.

Main Methods:

• Integrated elastic net-based selection into graph convolution to create GeC.
• Formulated GeC within a regularization framework, enabling a self-supervised update rule.
• Developed a new Graph Elastic Convolutional Network (GeCN) utilizing the GeC operation.

Main Results:

• GeC allows nodes to adaptively select optimal neighbors for feature aggregation.
• The proposed GeCN demonstrates effectiveness and robustness in experimental results.
• The self-supervised implementation of GeC simplifies its application.

Conclusions:

• The proposed GeC operation effectively addresses limitations of full neighborhood aggregation in GCNs.
• GeCN offers a robust and efficient approach to graph representation and learning.
• Adaptive neighbor selection is a promising direction for advancing GCNs.