DVGMAE: Self-Supervised Dynamic Variational Graph Masked Autoencoder.

Area of Science:

• Machine Learning
• Graph Neural Networks
• Artificial Intelligence

Background:

• Contrastive self-supervised learning (SSL) for dynamic graphs often requires extensive data augmentation and complex training techniques.
• Generative SSL, particularly masked autoencoders (MAEs), shows promise in overcoming these limitations but is underexplored in dynamic graph contexts.
• Key challenges include developing effective masking strategies and decoders that preserve temporal dependencies in dynamic graphs.

Purpose of the Study:

• To propose DVGMAE, a novel dynamic variational graph masked autoencoder.
• To address the challenges of masking strategy design and temporal dependency retention in dynamic graph MAE.
• To capture evolving behaviors and topological features in dynamic graphs effectively.

Main Methods:

• Introduced a temporal-aware masking strategy for dynamic graph edges, informed by historical mask data to reduce bias.
• Designed a globally enhanced decoder to reconstruct both temporal and spatial information from perturbed graph snapshots.
• Developed a variational graph masked autoencoder framework tailored for dynamic graph analysis.

Main Results:

• DVGMAE demonstrates superior performance compared to existing state-of-the-art methods.
• The proposed temporal-aware masking strategy effectively mitigates masking bias in dynamic graphs.
• The enhanced decoder successfully recovers crucial temporal and spatial graph information.

Conclusions:

• DVGMAE offers a robust generative SSL approach for dynamic graphs, overcoming limitations of previous methods.
• The model effectively captures dynamic graph evolution and topological structures.
• DVGMAE represents a significant advancement in generative SSL for dynamic graph representation learning.