Please join us听online听for a talk by听,听Assistant Professor听of Statistics at Purdue听University. She will听辫谤别蝉别苍迟听鈥淔ederated Learning for Feature Mismatch鈥� on Wednesday, September 23, at 1:30pm Central Time. For access to this seminar, contact Cierra Streeter.
Dr. Xue develops and analyzes advanced statistical methodology at the intersection of data integration, mobile health, missing data, high-dimensional inference, and statistical genetics. She received her PhD in Statistics from the University of Illinois Urbana鈥揅hampaign in 2019, followed by postdoctoral work at the University of Pennsylvania.听
Abstract听
We consider a framework that involves multiple sites with privacy constraints and heterogeneity, where different sites collect data from independent subjects, and site-specific, mismatched feature sets (e.g., hospitals with differing equipment and workflows). Most existing methods assume either aligned covariates or overlapping听cohorts, and听thus are inapplicable when both samples are听disjoint听and the covariates are mismatched. We propose Fed-PATE, a three-stage framework: local estimation at each site using its own observed features; a cross-site projection aggregation that recovers a consensus full-feature coefficients among similar sites; and a penalized transfer estimator that feeds the aggregated information back to each site. To accommodate heterogeneity and prevent negative transfer, we introduce a weighting scheme for model selection that adapts to inter-site similarity. Our proposed approach shows a lower prediction risk in theoretical analyses. Simulations show significant reductions in听prediction听MSE and robustness to inter-site heterogeneity. In a real-world application to Serum Creatinine (SCr) prediction using EHR variables, Fed-PATE improves performance over existing approaches for听the majority of听sites.听