Spatio-Temporal Graph Unlearning

The introduction of CallosumNet marks a pivotal advancement in the field of spatio-temporal graph modeling, particularly in light of stringent privacy regulations like GDPR and CCPA. These regulations necessitate the complete unlearning of unauthorized data, a challenge that existing methods struggle to meet due to their design for static graphs and localized data removal. CallosumNet, inspired by the brain's corpus callosum, employs innovative techniques such as Enhanced Subgraph Construction (ESC) and Global Ganglion Bridging (GGB) to efficiently erase data without the extensive costs associated with full model retraining. Empirical results from four diverse real-world datasets support the effectiveness of this framework, highlighting its potential to transform how sensitive data is managed in complex dynamic processes.