Abstract

In this talk, we begin by highlighting the role of spatial temporal graph neural networks (STGNNs) as effective models for capturing complex dependencies that evolve across both network structure and time. Although these models achieve strong predictive performance, they share a common limitation with many deep learning approaches: they provide only point-wise estimates and are often overconfident. This creates significant risks in real-world applications where operators require reliable assessments of uncertainty to support safe and informed decision-making.

To address this need, post-hoc uncertainty quantification methods have gained increasing attention. Conformal prediction is particularly appealing because it offers prediction intervals with finite sample guarantees without assuming a specific data distribution. However, standard conformal prediction depends on the exchangeability assumption. This assumption is violated in time series and spatial temporal settings, where data points are highly correlated and distributions evolve over time. As a result, vanilla conformal prediction does not provide valid coverage when applied directly to STGNN forecasts.

To overcome this challenge, we introduce a new method for uncertainty quantification in spatial-temporal data. The key idea is to use in-context learning for tabular foundation model to adapt nonconformity scores to local temporal conditions, which helps reduce the impact of exchangeability violations. This approach preserves the flexibility of STGN Ns while enabling conformal prediction to produce calibrated intervals in dynamic environments.

We apply our method to forecasting tasks in district heating networks using smart meter data. Our results show that our proposed method produces well-calibrated prediction intervals, remains robust under temporal distribution shifts, and outperforms existing conformal prediction baseline. This demonstrates that in-context adaptation provides a promising direction for reliable uncertainty quantification in spatial temporal graph learning.