Revisiting Deformable Convolution on Graphs: Large-Range Modeling and Robustness.

Area of Science:

• Graph Neural Networks
• Machine Learning
• Computer Science

Background:

• Traditional Graph Convolutional Networks (GCNs) use fixed receptive fields, limiting their ability to capture long-range dependencies.
• This limitation makes GCNs susceptible to graph attacks and noise, hindering their performance on complex graph data.

Purpose of the Study:

• To address the limitations of traditional GCNs by proposing a novel deformable graph convolution method.
• To enhance the representation power and robustness of graph learning models.

Main Methods:

• Introduced Neighborhood-Deformable Graph Convolution (NDGC), a novel deformable graph convolution technique.
• NDGC utilizes virtual neighbors with offsetting and interpolation to create a larger, deformable receptive field.
• Message aggregation is performed on these deformable virtual neighbors for enhanced robustness.

Main Results:

• NDGC effectively captures long-range dependencies between distant nodes in graph structures.
• The proposed method demonstrates increased robustness against graph attacks and noise compared to traditional GCNs.
• NDGC integrates seamlessly with existing GCNs, creating deformable variants.

Conclusions:

• NDGC offers a significant advancement in graph representation learning by overcoming the limitations of fixed receptive fields.
• The method's effectiveness and advantages are validated across various graph learning tasks.
• NDGC provides a generalizable framework for developing more powerful and resilient graph convolutional networks.