Laplacian regularized stratified models

Details

Supervisors

Pircalabelu, Eugen

Faculty

Faculté des sciences

Degree label

Master [120] en science des données, orientation statistique, à finalité spécialisée

Abstract

This thesis investigates Laplacian Regularized Stratified Models (LRSMs) and their extensions. These methods aim to improve parametric models trained on distinct subgroups within a dataset. Indeed, while stratification enables more specialized predictions, it also fragments the data available for training, which may lead to unreliable parameter estimates. To address this issue, LRSMs introduce a graph based regularization which penalizes the pairwise difference between group parameters. This mechanism enables each individual model to "borrow strength" from the estimates of its neighbours to improve its own.