This week’s seminar brings another thought-provoking discussion! Join us on February 11th to hear from CTML GSR Mingxun (Michael) Wang presenting his talk on "Highly Adaptive Principal Component Regression: Fast HAL/HAR via Outcome-Blind Kernel PCA." The seminar will take place at 12:00 PM in Berkeley Way West, 5th Floor, Room 5401.
Abstract: The Highly Adaptive Lasso (HAL) has strong rate guarantees under minimal smoothness assumptions, but can be computationally prohibitive in moderate to high dimensions due to its enormous basis expansion. We introduce PCHAL and PCHAR, which perform outcome-blind dimension reduction by projecting the highly adaptive kernel onto its leading principal components, yielding simple closed-form ridge solutions and a lasso solution that reduces to soft-thresholding in an orthogonal score space. The resulting estimators substantially accelerate fitting and cross-validation while matching the empirical predictive performance of HAL/HAR, and are implemented in the hapc R package.