Spatial Entropy Speeds Spatiotemporal Graph Unlearning for Privacy

Spatiotemporal graphs are crucial for modeling dynamic systems like traffic or weather, but complying with data privacy regulations such as GDPR poses a significant challenge. Erasing specific data points from these complex models typically requires costly full-graph retraining due to the global propagation of information. Researchers have introduced IsleNet, a novel framework designed to make this process both exact and efficient. IsleNet partitions the spatiotemporal graph into coherent subgraphs using spatial entropy, connecting them with lightweight virtual edges. When an unlearning request is made, only the relevant subgraph encoder and virtual-edge layer need to be retrained. This targeted approach ensures complete data removal with a substantial reduction in computational cost. Experimental results demonstrate that IsleNet achieves accuracy comparable to full-graph models while speeding up the unlearning process by an order of magnitude.