Temporal Network Embedding for Link Prediction via VAE Joint Attention Mechanism.

Area of Science:

• Computer Science
• Data Science
• Network Science

Background:

• Temporal networks exhibit dynamic topological structures over time.
• Existing temporal network embedding (TNE) methods struggle with temporal dependencies, topological changes, and computational redundancy.
• Challenges include describing temporal dependence across snapshots, indicating topological changes in latent space, and avoiding redundant computations.

Purpose of the Study:

• To propose a novel temporal network embedding method (TVAE) based on the Variational Autoencoder (VAE) framework.
• To capture the evolution of temporal networks for effective link prediction.
• To overcome limitations of existing TNE methods regarding temporal dependence, topological change representation, and computational efficiency.

Main Methods:

• Developed a Variational Autoencoder (VAE)-based framework for temporal network embedding (TVAE).
• Integrated self-attention mechanisms and recurrent neural networks to update node representations and maintain temporal dependencies.
• Employed parameter inheritance for efficient updates and scalability to large networks.

Main Results:

• TVAE generates low-dimensional node embeddings that preserve dynamic nonlinear features of temporal networks.
• The model effectively captures temporal dependencies and network evolution.
• Experimental results show TVAE outperforms baseline methods in performance and reduces time cost.

Conclusions:

• TVAE offers a superior approach for temporal network embedding and link prediction.
• The method is computationally efficient and scalable for large-scale dynamic networks.
• TVAE successfully addresses key challenges in analyzing evolving network structures.