Time-Aware Graph Learning for Link Prediction on Temporal Networks.

Area of Science:

• Graph Representation Learning
• Network Science
• Machine Learning

Background:

• Link prediction in temporal networks is crucial for understanding dynamic graph evolution.
• Existing methods struggle with attribute deficiency, distance estimation limitations, and inadequate modeling of dynamic node correlations by graph neural networks (GNNs).

Purpose of the Study:

• To propose a novel time-aware graph (TAG) learning method for accurate link prediction in temporal networks.
• To address the limitations of existing GNN-based approaches in capturing dynamic node correlations.

Main Methods:

• Theoretical causal analysis to establish conditions for temporal graph representation learning.
• An edge-dropping (ED) module and recent neighbor sampling (RNS) to model recent dynamic node correlations.
• Contrastive learning for self-supervision to preserve long-term stable node correlations.

Main Results:

• TAG achieves state-of-the-art performance on four public temporal network datasets (MathOverflow, StackOverflow, AskUbuntu, SuperUser).
• Demonstrated improvements in average precision (AP) and area under the ROC curve (AUC).
• TAG ensures high computational efficiency through a lightweight temporal graph representation.

Conclusions:

• The proposed TAG method effectively predicts future edges in temporal networks by accurately modeling dynamic node correlations.
• TAG offers a practical and computationally efficient solution for real-world temporal network link prediction applications.