A0972
Title: On error distribution in nonparametric Bayesian functional clustering
Authors: Tomoya Wakayama - RIKEN AIP (Japan) [presenting]
Fumiya Iwashige - Graduate School of Advanced Science and Engineering, Hiroshima University (Japan)
Shintaro Hashimoto - Hiroshima University (Japan)
Shonosuke Sugasawa - Keio University (Japan)
Abstract: Nonparametric Bayesian functional clustering often returns too many clusters empirically. Existing work blames partition priors and proposes heavier-tailed processes, yet most models still assume independent observation noise. It is proven that when functions are observed on a dense grid, this misspecification forces the posterior to favor the singleton partition. In light of this phenomenon, it is discussed what sort of modeling is necessary for functional clustering.
