Presentation 2-3, 14:20~14:45.
Speaker
Junseok Lee (Seoul National University)
Title
Constructing Conformal Predictive Distributions via Kernel-Localized Residuals
Abstract
We present a method that uses conformal prediction to construct a predictive distribution for each new input. First, residuals on calibration data are used to form an empirical distribution instead of a global normalized-residual distribution across inputs. Second, kernel weights localize this empirical distribution around the test point so that nearby data count more. Sampling from the localized distribution yields the predictive distribution, and the prediction interval is derived directly from it. Formed locally, the distribution captures complex shapes, mitigates location bias, and adapts to changes in support, while preserving marginal coverage. The method keeps a clear split between training and calibration, works with a wide range of base models, and does not require density modeling. Random projections define locality in high dimensions. On synthetic and real data, performance is comparable in low dimensions and clearly better in high dimensions, where the error distribution varies locally in complex ways.