Multiscale Subgraph Adversarial Contrastive Learning | JoVE Visualize

Area of Science:

Artificial Intelligence
Machine Learning
Graph Neural Networks

Background:

Graph contrastive learning (GCL) is a prominent self-supervised learning paradigm for graph representation.
Existing GCL methods assume augmented views from the same graph are positive pairs, which may fail for complex, multiscale graphs.
The semantic consistency of augmented graph views is crucial but often overlooked.

Purpose of the Study:

To address the limitations of standard GCL in handling multiscale graph structures.
To propose a novel GCL method that captures fine-grained semantic information by considering multiscale graph properties.
To enhance the generalization performance of GCL models through advanced techniques.

Main Methods:

Developed a multiscale subgraph contrastive learning (MSSGCL) method using subgraph sampling for global and local views.
Constructed multiple contrastive relationships based on semantic associations at different scales.
Introduced MSSGCL++ with an asymmetric structure and adversarial training for improved generalization.
Optimized a min-max saddle point problem and employed a "free" strategy for faster training.

Main Results:

Experimental analysis revealed that semantic information consistency is vital and depends on graph multiscale structure.
MSSGCL effectively characterizes fine-grained semantic information by leveraging multiscale subgraph views.
MSSGCL++ demonstrated improved generalization performance compared to the base MSSGCL model.
The proposed methods achieved significant improvements over state-of-the-art approaches on 16 real-world graph classification datasets.

Conclusions:

Standard GCL assumptions do not always hold for complex, multiscale graphs.
Multiscale subgraph contrastive learning offers a more robust approach to graph representation learning.
The proposed MSSGCL and MSSGCL++ methods significantly advance the field of self-supervised graph learning.
The findings highlight the importance of considering graph structure at multiple scales for effective representation learning.