Temporal Network Embedding Enhanced With Long-Range Dynamics and Self-Supervised Learning.

Area of Science:

• Computer Science
• Network Science
• Data Mining

Background:

• Temporal network embedding (TNE) is crucial for knowledge discovery and reasoning on dynamic networks.
• Existing TNE methods struggle to capture long-distance dynamics, limiting the exploration of multihop topological associations.
• This limitation hinders the understanding of complex relationships in evolving networks.

Purpose of the Study:

• To propose LongTNE, a novel method for temporal network embedding that captures long-range vertex dynamics.
• To enable TNE methods to effectively capture high-order proximity (HP) in temporal networks.
• To improve the performance of network mining tasks by addressing limitations in capturing temporal network structures.

Main Methods:

• LongTNE utilizes graph self-supervised learning (Graph SSL) to optimize deep link establishment probabilities within network snapshots.
• An accumulated forward update (AFU) module is introduced to analyze global temporal evolution across multiple network snapshots.
• The method focuses on learning long-range dynamics of vertices for enhanced network representation.

Main Results:

• Empirical results on six temporal networks demonstrate LongTNE's state-of-the-art performance in network mining tasks.
• LongTNE effectively captures high-order proximity, overcoming limitations of existing TNE methods.
• The proposed method shows significant improvements in preserving network structures and temporal properties.

Conclusions:

• LongTNE offers a powerful approach to temporal network embedding by capturing long-range dynamics and high-order proximity.
• The method achieves superior performance on various network mining tasks.
• LongTNE is adaptable and can be readily extended to enhance existing temporal network embedding techniques.