RAFT: Redundancy-Aware Input Filtering for Temporal Graph Neural Network Training

Key points

• As graphs grow rapidly in modern applications, training requires processing an increasing number of events; meanwhile, each event can update node memories used by later events, creating dependencies that constrain parallel execution and limit the benefits of additional computing resources.1
• We observe that targets involving stable node memories tend to exhibit lower prediction loss, while filtering historical events associated with stable memories can retain similar training-loss trajectories in several evaluated settings.1
• Building on this opportunity, we propose RAFT, a redundancy-aware input-filtering framework for TGNN training.1
• RAFT tracks memory changes during training and uses them to guide two filters: a root-event filter removes selected prediction targets before training, and a support-event filter excludes selected historical context during sampling.1

Sources (1)

• [1]RAFT: Redundancy-Aware Input Filtering for Temporal Graph Neural Network Training
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 17, 09:26 PM

  Temporal graph neural networks (TGNNs) have shown strong performance on dynamic graph learning tasks by continuously updating node memories from evolving interactions (i.e., events) and leveraging these memories for downstream predictions.

  As graphs grow rapidly in modern applications, training requires processing an increasing number of events; meanwhile, each event can update node memories used by later events, creating dependencies that constrain parallel execution and limit the benefits of additional computing resources.

Extractive summary: sentences quoted from the sources.