About the Talk
Conformal prediction turns any prediction model into a prediction interval with a coverage guarantee. It sets the model aside, measures how badly it missed on held-out data, and allows a new prediction to miss by about as much. The guarantee holds in finite samples and for any model, provided the new observation is no more unusual than the held-out ones.
Spatial data appears to violate this, since nearby observations are correlated. We show that correlation alone is not the problem. What matters is where the observations are: scattered locations pose no difficulty, while clustered locations, or a prediction far from all of them, do. Existing methods address one case or the other, and the method that handles the harder case does so by discarding most of the held-out data.
We propose a single method for both. Instead of discarding correlated observations, we transform them so that each contributes only the information the others do not already carry. We establish exact finite-sample coverage when the spatial model is correct, bound the loss when it is not, and derive a quantity, computable from the data, that predicts how much coverage a given region will lose before any is observed.
About the Speaker
Arkajyoti (Arka) Saha is an Assistant Professor of Statistics in the Donald Bren School of Information and Computer Sciences at the University of California, Irvine. He integrates statistical modeling with artificial intelligence and machine learning to develop methods that account for dependence structures in data. His work bridges classical statistical theory with modern AI/ML to create scalable, interpretable, and scientifically grounded learning frameworks.
Before joining UC Irvine, Arkajyoti was a Postdoctoral Fellow in the Department of Statistics at the University of Washington, where he worked with Daniela Witten and Jacob Bien. He received his PhD in Biostatistics from the Johns Hopkins Bloomberg School of Public Health, where he was co-advised by Nilanjan Chatterjee and Abhirup Datta.
