Multi-angle information aggregation for inductive temporal graph embedding

Area of Science:

• Computer Science
• Machine Learning
• Graph Theory

Background:

• Graph embedding represents large graphs in low-dimensional spaces.
• Transductive learning methods are limited for dynamic temporal graphs.
• Temporal graphs require methods that handle continuous changes.

Purpose of the Study:

• Propose an inductive temporal graph embedding method (MIAN).
• Address limitations of transductive approaches for evolving graphs.
• Develop a model for effective node embedding in dynamic networks.

Main Methods:

• MIAN employs Multi-angle Information Aggregation Network.
• Information is aggregated from neighborhood, temporal, and environment angles.
• An improved gated recurrent unit (GRU) module combines aggregated information.

Main Results:

• MIAN demonstrates superior performance on real-world datasets.
• Experimental results show effectiveness across diverse tasks.
• The proposed method outperforms state-of-the-art baselines.

Conclusions:

• MIAN provides an effective inductive approach for temporal graph embedding.
• The multi-angle information aggregation is key to MIAN's success.
• This method advances dynamic graph representation learning.