Temporal Graph Pattern Machine
A new AI framework uses 'interaction patches' and self-supervised learning to model how networks change over time.
Researchers Yijun Ma, Zehong Wang, Weixiang Sun, and Yanfang Ye developed the Temporal Graph Pattern Machine (TGPM), a foundation model for dynamic networks. It uses temporally-biased random walks to create 'interaction patches' and a Transformer backbone to capture long-range dependencies. Pre-trained with masked token and next-time prediction tasks, it achieves state-of-the-art results in link prediction and shows strong cross-domain transferability, outperforming task-specific models.
Why It Matters
This enables more accurate forecasting in social networks, financial transactions, and logistics by understanding fundamental, reusable patterns of change.