Spatio-Temporal Graph Unlearning
Qiming Guo, Wenbo Sun, Wenlu Wang
TL;DR
This paper introduces CallosumNet, a novel framework for complete unlearning in spatio-temporal graphs, enabling efficient removal of data while maintaining model performance, inspired by brain structure.
Contribution
We propose CallosumNet, a divide-and-conquer approach with novel techniques ESC and GGB for effective unlearning in complex spatio-temporal graphs.
Findings
Achieves complete unlearning with only 1-2% MAE loss.
Outperforms existing baselines significantly.
Effective ablation results confirm the techniques' contributions.
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…
Peer Reviews
Decision·ICLR 2026 Conference Withdrawn Submission
The overall framework is clearly described, and the motivation is easy to follow. The authors conduct comprehensive experiments across multiple datasets and baselines, showing that their method achieves better performance than existing approaches while preserving accuracy close to the gold model. The results demonstrate consistent improvements, and the ablation study helps clarify the role of each component.
The related work section, while broad, lacks a deep comparison with recent developments in efficient retraining or federated unlearning. Experimental baselines are not comprehensive; several more recent or specialized spatio-temporal unlearning methods are omitted. The analysis of results remains mostly descriptive without strong theoretical or empirical evidence for the claims and there is no practical validation of privacy compliance.
- Originality: This is the first principled framework for exact unlearning in spatio-temporal graphs, a domain where influence propagates across both space and time. The biological inspiration (corpus callosum) is novel and well-justified. The idea of virtual ganglion edges and meta-graph fusion is creative and technically sound. - Technical Quality: The paper is rigorous in both design and theory. The authors provide formal guarantees for unlearning exactness, prediction error bounds, and mode
- Limited Theoretical Analysis of Approximation Error: While the paper provides unlearning exactness, it lacks a fine-grained analysis of how approximations in ESC and GGB affect long-term prediction stability. For example, how does the meta-graph fusion affect error propagation over time? A perturbation analysis or error bound under cumulative unlearning would strengthen the contribution. - Scalability to Larger Graphs: Although the method is tested on graphs with up to 3,220 nodes, urban-scal
1. This paper focuses on a unique problem. 2. Experiments demonstrate the effectiveness of the model.
- One of the biggest flaws of the paper is that I cannot see the necessity of unlearning in spatio-temporal forecasting. The authors mention privacy issues, but in reality, most spatio-temporal datasets are open-source and do not involve personal information. The statement, "As a result, ensuring compliance with these privacy requirements often requires retraining the entire spatio-temporal graph model to preserve privacy for individual nodes, a process that, while essential, introduces addition
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsAdvanced Graph Neural Networks · Machine Learning in Healthcare · Graph Theory and Algorithms