Fine-Grained Interactive Transformers for Continuous Dynamic Link Prediction.

Area of Science:

• Network Science
• Machine Learning
• Data Mining

Background:

• Dynamic link prediction (DLP) is crucial for understanding evolving systems.
• Existing DLP methods struggle with independent node interaction modeling and fine-grained latent interaction characterization.

Purpose of the Study:

• To propose FineFormer, a novel framework for accurate dynamic link prediction.
• To address limitations in modeling intra-node and inter-node temporal dependencies.

Main Methods:

• FineFormer utilizes alternating self-attention and cross-attention mechanisms.
• Layer-wise contrastive learning is employed to refine node representations.
• Self-attention models temporal-spatial dynamics within individual node sequences.
• Cross-attention captures complex interactions across node pair sequences.

Main Results:

• FineFormer consistently outperforms state-of-the-art baselines on five diverse DLP datasets.
• The framework excels at capturing complex, fine-grained interactions in continuous-time dynamic networks.
• Experimental results validate the effectiveness of the proposed attention and contrastive learning strategies.

Conclusions:

• FineFormer offers a significant advancement in dynamic link prediction.
• The model's ability to capture fine-grained temporal dependencies improves prediction accuracy.
• This framework provides a robust solution for analyzing and forecasting evolving network structures.