ScaDyG: A New Paradigm for Large-Scale Dynamic Graph Learning.

Area of Science:

• Machine Learning
• Graph Neural Networks
• Dynamic Systems

Background:

• Dynamic graphs (DGs) model time-evolving relationships, crucial for many real-world applications.
• Existing dynamic graph neural networks (DGNNs) face scalability challenges due to historical data growth.
• Industry applications require efficient encoding of DGs for downstream tasks.

Purpose of the Study:

• To propose ScaDyG, a novel time-aware scalable learning paradigm for dynamic graphs.
• To address the scalability limitations of traditional DGNNs.
• To improve efficiency and performance in DG encoding for downstream tasks.

Main Methods:

• Time-aware topology reformulation (TTR): Segments historical interactions into time steps for weight-free, time-aware propagation.
• Dynamic temporal encoding (DTE): Integrates fine-grained temporal encoding using exponential functions.
• Hypernetwork-driven message aggregation: Employs a hypernetwork for adaptive temporal fusion of node representations.

Main Results:

• ScaDyG demonstrates comparable or superior performance to state-of-the-art (SOTA) methods on 12 datasets.
• Achieved strong results in both node-level and link-level downstream tasks.
• Exhibited fewer learnable parameters and higher computational efficiency compared to existing methods.

Conclusions:

• ScaDyG offers an effective and efficient solution for learning on dynamic graphs.
• The proposed methods successfully address scalability issues in DGNNs.
• The approach provides a promising direction for real-world DG applications.