Learning dynamic graph representations through timespan view contrasts.

Area of Science:

• Graph Representation Learning
• Dynamic Graph Analysis
• Machine Learning

Background:

• Unsupervised graph representation often overlooks temporal dynamics in real-world data.
• Existing methods rely on static graph properties, neglecting edge timestamps.
• Modeling temporal evolution in dynamic graphs remains a challenge.

Purpose of the Study:

• To develop an elegant framework for modeling temporal evolution on dynamic graphs.
• To introduce and leverage the inductive bias of temporal translation invariance.
• To enhance dynamic graph representation learning and anomaly detection.

Main Methods:

• Proposed CLDG framework utilizing contrastive learning across different timespans.
• Introduced temporal translation invariance as a key inductive bias.
• CLDG++ incorporates graph diffusion for global correlations and multi-scale contrastive objectives.

Main Results:

• CLDG and CLDG++ demonstrate strong performance in node classification and dynamic graph anomaly detection.
• CLDG reduces time and space complexity by implicitly using temporal cues.
• The proposed methods effectively identify anomalies in dynamic graphs across various domains.

Conclusions:

• CLDG offers an efficient and effective approach to dynamic graph representation learning.
• Temporal translation invariance is a valuable bias for modeling dynamic graph evolution.
• The framework shows significant potential for applications in finance, cybersecurity, and healthcare.