SG-JEPA spiking architecture scales dynamic graph learning to 13M edges

Dynamic graph learning—crucial for fraud detection and recommendations—has long struggled with scalability of self-supervised methods. Existing approaches rely on heavy machinery like edge-level reconstruction or graph contrastive learning, which blow up computational costs on large graphs. A new paper from Sun Yat-sen University introduces SG-JEPA, a spiking neural network (SNN) architecture that reimagines self-supervision for dynamic graphs. Instead of reconstructing edges or augmenting the graph, SG-JEPA splits nodes into context and target sets along the temporal axis, then learns embeddings that predict one another using spatial-temporal information. The key innovation is encoding sequential inputs into coarse-to-fine spike count embeddings via spiking neurons. This allows the model to adapt computational load to downstream task constraints—a major advantage for production systems.

Experimental results confirm SG-JEPA scales to a dynamic graph with 13 million edges while achieving competitive or superior performance on node classification compared to discriminative baselines. More importantly, it sidesteps the expensive pipelines of prior self-supervised methods: no negative sampling, no graph augmentations, no edge-level reconstruction. Training efficiency and memory scalability are both significantly improved. For professionals building real-time graph-based systems—social networks, financial fraud detection, or recommender engines—SG-JEPA offers a practical path to self-supervised learning at scale without the usual resource tax.