Temporal graph neural networks (TGNNs) are designed for dynamic graph learning, continuously updating node memories from evolving interactions, or events. However, not all events are equally informative, and feeding redundant inputs into the model can waste computation and slow down training.
The arXiv paper introduces RAFT, a redundancy-aware input filtering mechanism that aims to identify and filter less useful events before they reach the model. By reducing the volume of redundant data, RAFT could make TGNN training more efficient while preserving the model's ability to learn from meaningful interactions.
The abstract is brief, so specific experimental results are not available in the source. The contribution lies in the concept of redundancy-aware filtering for TGNN training, addressing a practical bottleneck in dynamic graph learning.