Dynamic graph representation learning with disentangled information bottleneck.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Graph Neural Networks

Background:

• Dynamic graph representation learning is crucial but existing methods often entangle temporal properties.
• Holistic approaches overlook the dichotomy of time-invariant and time-varying properties in dynamic graphs.
• This entanglement can lead to suboptimal performance in downstream applications.

Purpose of the Study:

• To propose a novel method for learning macro-disentangled dynamic graph representations.
• To address the limitations of existing methods in handling diverse temporal dependencies.
• To enhance the performance of dynamic graph representations in various tasks.

Main Methods:

• Introduced Disentangled Dynamic Graph Information Bottleneck (DDGIB), a novel dynamic graph representation learning method.
• Leveraged Information Bottleneck theory for macro-disentanglement.
• Explicitly embedded dynamic graphs into separate time-invariant and time-varying representation spaces.

Main Results:

• DDGIB successfully disentangles time-invariant and time-varying properties of dynamic graphs.
• Theoretical proofs confirm the sufficiency and macro disentanglement of the DDGIB method.
• Extensive experiments demonstrated the superiority of DDGIB across various datasets and tasks.

Conclusions:

• DDGIB offers a powerful approach to dynamic graph representation learning by disentangling temporal properties.
• The method achieves sufficient representations for downstream tasks.
• DDGIB enhances model performance by capturing distinct temporal dynamics effectively.