CapMax: A Framework for Dynamic Network Representation Learning From the View of Multiuser Communication.

Area of Science:

• Network Science
• Information Theory
• Machine Learning

Background:

• Dynamic networks possess time-varying topology and node attributes, posing challenges for representation learning.
• Existing methods often rely on specific backbone structures, limiting their applicability.
• Mutual Information (MI) is commonly used but may not fully capture information in dynamic systems.

Purpose of the Study:

• To propose a novel framework, Channel Capacity Maximization (CapMax), for learning informative representations of dynamic networks.
• To adapt principles from multiuser communication theory to model network representation learning.
• To enhance the capture of time-varying information compared to traditional MI-based approaches.

Main Methods:

• Developed CapMax, a modified mutual information maximization (InfoMax) framework.
• Treated the representation model as a multiaccess communication channel with memory and feedback.
• Utilized Graph Neural Networks (GNNs) and Recurrent Neural Networks (RNNs) to estimate channel capacity via Directed Information (DI).

Main Results:

• Demonstrated theoretically that Directed Information (DI) is superior to Mutual Information (MI) for capturing useful information under mild conditions.
• Empirically validated CapMax on multiple real-world dynamic network datasets.
• Showcased CapMax's outperformance across various backbone models in link detection and prediction tasks.

Conclusions:

• CapMax effectively learns informative representations for dynamic networks by modeling them as communication channels.
• The proposed DI-based approach offers a more comprehensive measure of information in dynamic network contexts.
• The framework's flexibility and effectiveness are confirmed by its superior performance on benchmark tasks.