BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Conformalizing Spatial-Temporal Graph Neural Networks with In-Cont
 ext Learning: Case Studies in District Heating Networks - Keivan Faghih Ni
 resi\, PhD student\, Swiss Federal Institute of Technology in Lausanne (EP
 FL)
DTSTART:20251203T160000Z
DTEND:20251203T170000Z
UID:TALK241441@talks.cam.ac.uk
CONTACT:46601
DESCRIPTION:In this talk\, we begin by highlighting the role of spatial te
 mporal graph neural networks (STGNNs) as effective models for capturing co
 mplex dependencies that evolve across both network structure and time. Alt
 hough these models achieve strong predictive performance\, they share a co
 mmon limitation with many deep learning approaches: they provide only poin
 t-wise estimates and are often overconfident. This creates significant ris
 ks in real-world applications where operators require reliable assessments
  of uncertainty to support safe and informed decision-making.\n \nTo addre
 ss this need\, post-hoc uncertainty quantification methods have gained inc
 reasing attention. Conformal prediction is particularly appealing because 
 it offers prediction intervals with finite sample guarantees without assum
 ing 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 c
 orrelated and distributions evolve over time. As a result\, vanilla confor
 mal prediction does not provide valid coverage when applied directly to ST
 GNN forecasts.\n\nTo overcome this challenge\, we introduce a new method f
 or uncertainty quantification in spatial-temporal data. The key idea is to
  use in-context learning for tabular foundation model to adapt nonconformi
 ty scores to local temporal conditions\, which helps reduce the impact of 
 exchangeability violations. This approach preserves the flexibility of STG
 NNs while enabling conformal prediction to produce calibrated intervals in
  dynamic environments.\n\nWe apply our method to forecasting tasks in dist
 rict heating networks using smart meter data. Our results show that our pr
 oposed method produces well-calibrated prediction intervals\, remains robu
 st under temporal distribution shifts\, and outperforms existing conformal
  prediction baseline. This demonstrates that in-context adaptation provide
 s a promising direction for reliable uncertainty quantification in spatial
  temporal graph learning.
LOCATION:LR3B\, Inglis Building\, CUED.
END:VEVENT
END:VCALENDAR