Spiking Neural Networks Boost Large Dynamic Graph Learning.

This research introduces SG-JEPA (Joint Spiking Embedding Predictive Architecture), a novel framework for learning from large-scale dynamic graphs, which are common in applications like fraud detection and recommender systems. Existing self-supervised methods for dynamic graphs often incur substantial computational overhead due to complex edge-level reconstruction objectives and tailored graph augmentation strategies. SG-JEPA addresses these limitations by adopting a simpler, more efficient approach. It partitions nodes into context and target sets along the temporal dimension, learning embeddings that are mutually predictive through additional spatial-temporal information. A key innovation is the use of spiking neurons, which encode sequential inputs into coarse-to-fine spike count embeddings, allowing the architecture to adapt to varying computational constraints of downstream tasks. This results in competitive or superior performance on node classification while achieving significantly better training efficiency and memory scalability compared to prior self-supervised dynamic graph baselines.