Self-Supervised Continuous Dynamic Graph Representation Learning via Hawkes Processes.

Area of Science:

• Machine Learning
• Graph Neural Networks
• Temporal Data Analysis

Background:

• Dynamic graph representation learning (DGRL) is crucial for real-world applications but often requires costly labeled data.
• Existing DGRL methods struggle to capture temporal interactions and event persistence, leading to suboptimal representations.
• The cumulative impact of past events in dynamic graphs can be complex and sequence-dependent.

Purpose of the Study:

• To introduce a novel self-supervised framework for dynamic graph representation learning (DGRL).
• To enhance temporal representations by accurately modeling event interactions, including excitation and inhibition.
• To develop an efficient and versatile DGRL method that generalizes to various downstream tasks without task-specific encoders.

Main Methods:

• Proposed self-supervised dynamic graph representation learning (SDGRL) framework based on Hawkes processes.
• Leveraged Hawkes processes to model multi-perspective event interactions, capturing temporal dependencies.
• Demonstrated theoretical properties including stationarity and convergence of the SDGRL method.

Main Results:

• SDGRL achieves optimal performance across eight real-world datasets, outperforming nine existing methods.
• On the Tmall dataset, SDGRL improved inductive dynamic link prediction accuracy by 28.73% compared to DyGFormer.
• SDGRL significantly reduced computational costs, decreasing time by 1.29x and memory usage by 15.33x.

Conclusions:

• SDGRL offers an effective self-supervised approach for DGRL, eliminating the need for labeled data.
• The framework accurately captures complex temporal event interactions, leading to superior node embeddings.
• SDGRL provides a computationally efficient and versatile solution for dynamic graph analysis.