A0676
Title: Parametric bootstrap evaluation of unsupervised statistical learning and applications
Authors: Berthold Lausen - University of Essex (United Kingdom) [presenting]
Abstract: Ultrametrics and additive tree metrics are mathematical models of phylogenetic inference or hierarchical clustering of distance data, which can be seen as unsupervised statistical learning problems. An additive measurement error model for distance data was used previously to develop a three objects variance estimator which provides a point estimate of the variance parameter without estimating the overall phylogenetic tree as an ultrametric or additive tree. Estimating the unknown location parameter, ultrametric or dendrogram, the three-objects variance estimator is used to compute parametric bootstrap estimates of the probability to observe the estimated clusters. The approach is applied in the context of user segmentation based on online behavioural data and compared to other recent suggestions for the evaluation of unsupervised statistical learning.
