ConTIG: Continuous representation learning on temporal interaction graphs.

Area of Science:

• Graph Representation Learning
• Network Science
• Machine Learning

Background:

• Temporal Interaction Graphs (TIGs) are crucial for modeling dynamic networks in various applications.
• Existing TIG methods often fail to capture continuous embedding evolution or overlook temporal patterns, leading to suboptimal performance.

Purpose of the Study:

• To propose ConTIG, a novel two-module framework for representation learning on TIGs.
• To capture the continuous dynamic evolution of node embedding trajectories in TIGs.
• To improve the accuracy of dynamic network analysis tasks.

Main Methods:

• ConTIG utilizes a two-module framework incorporating continuous inference and self-attention mechanisms.
• The first module employs ordinary differential equations to learn node state trajectories from time-adjacent interactions.
• The second module uses self-attention to predict future embeddings by aggregating historical temporal interaction data.

Main Results:

• ConTIG demonstrates superior performance on temporal link prediction, node recommendation, and dynamic node classification tasks.
• Experiments were conducted on four datasets, comparing ConTIG against state-of-the-art baselines.
• The model shows particular effectiveness in predicting interactions over long intervals.

Conclusions:

• ConTIG effectively models continuous node embedding trajectories in dynamic networks.
• The proposed framework enhances the understanding and prediction capabilities for Temporal Interaction Graphs.
• ConTIG offers a significant advancement over existing methods for dynamic network analysis.