Spatio-Temporal Graph Unlearning

TL;DR

CallosumNet is proposed, a divide-and-conquer spatio-temporal graph unlearning framework inspired by the corpus callosum structure that facilitates communication between the brain's two hemispheres that achieves complete unlearning with only 1%-2% relative MAE loss compared to the gold model, significantly outperforming state-of-the-art baselines.

Abstract

Spatio-temporal graphs are widely used in modeling complex dynamic processes such as traffic forecasting, molecular dynamics, and healthcare monitoring. Recently, stringent privacy regulations such as GDPR and CCPA have introduced significant new challenges for existing spatio-temporal graph models, requiring complete unlearning of unauthorized data. Since each node in a spatio-temporal graph diffuses information globally across both spatial and temporal dimensions, existing unlearning methods primarily designed for static graphs and localized data removal cannot efficiently erase a single node without incurring costs nearly equivalent to full model retraining. Therefore, an effective approach for complete spatio-temporal graph unlearning is a pressing need. To address this, we propose CallosumNet, a divide-and-conquer spatio-temporal graph unlearning framework inspired by the corpus callosum structure that facilitates communication between the brain's two hemispheres. CallosumNet incorporates two novel techniques: (1) Enhanced Subgraph Construction (ESC), which adaptively constructs multiple localized subgraphs based on several factors, including biologically-inspired virtual ganglions; and (2) Global Ganglion Bridging (GGB), which reconstructs global spatio-temporal dependencies from these localized subgraphs, effectively restoring the full graph representation. Empirical results on four diverse real-world datasets show that CallosumNet achieves complete unlearning with only 1%-2% relative MAE loss compared to the gold model, significantly outperforming state-of-the-art baselines. Ablation studies verify the effectiveness of both proposed techniques.

Figures (4)

• Figure 1: Unlearning on a spatio-temporal graph. (a) A fully connected ST-Graph yields high accuracy; node size encodes impact, color encodes evolving features, and arrows denote spatio-temporal edges. (b) Red marks indicate users who revoke data-use consent; their nodes and incident edges must be erased. (c) Deleting raw records satisfies compliance yet leaves residual influence (faded arrows) inside the model. (d) Retraining after deletion purges influence but fragments the graph and distorts remaining node features ($v_1, v_2$, $v_3$), harming accuracy.
• Figure 2: The corpus callosum. A bridge of $\sim$2 $\times$ 10$^{8}$ axons connecting the two cerebral hemispheres. Although comprising only about 1% of each hemisphere's $\sim$20 billion cortical neurons, it provides ample bandwidth to synchronise bilateral neural activity.
• Figure 3: CallosumNet system construction. The original graph (a) is transformed into multiple enhanced local subgraphs (d) through ESC, and then the Global Ganglion Bridging (GGB) method adds ganglion nodes and identifies key nodes to construct the meta-graph. All the enhanced local subgraphs are trained into enhanced sub-models, with their weights frozen. These sub-models, along with the ganglion nodes and the global integration slot, are combined to form CallosumNet. After a small amount of training data updates the parameters, the entire CallosumNet can operate normally with prediction accuracy within 1%-2% of the original ST-Graph.
• Figure 4: CallosumNet unlearning process