Date/Time
Date(s) - 01/11/2022
3:30 pm - 5:00 pm

Title: The Good, the Bad, and the Asymmetric: Clustering with Finite Mixtures of SU Johnson Distributions

Speaker: Nikola Pocuca, McMaster University

Abstract: The presence of asymmetry in realistic datasets contrives the use of advanced techniques to handle skewness and kurtosis.  A considerable amount of literature has been developed over the years for such scenarios; ranging from transformations to heavy-tailed distributions. In recent years, these techniques have been adapted for clustering heterogeneous data that are considerable in both size, skewness, and dimension.  Most notably, the mixture model-based approach has shown great efficacy for clustering such high dimensional scenarios. In particular, the use of transformations within a model-based framework  has proven to be effective in dealing with skewed data. Despite the popularity of  transformation techniques, there is a general paucity within literature regarding the SU Johnson distribution.  An alternative to the popularized power transformation, the  SU distribution has been shown among contemporaries to have superior performance overall.  In this work, we develop a mixture model-based approach for clustering skewed data using finite mixtures of SU distributions. Additionally, we also develop an imputation method to handle missing data scenarios. Using realistic datasets, our method proves itself highly robust in the presence of heterogeneity, and asymmetry.

Date/Time: Tuesday November 1, 2022, 3:30 – 5:00 

Location: UH 112