Unsupervised Classification for Skewed and Mixed-Type Data

Abstract

Clustering, also known as unsupervised classification, is a foundational machine learning technique and is used to find underlying group structures in data. There are many well-established model-based techniques to analyze either categorical or continuous data in the clustering paradigm. However, there is a relative paucity of work for mixed-type data, especially mixed data where the continuous variables exhibit skewness and heavy tails. In this thesis, different methodologies and models are presented for analyzing asymmetric and mixed-typed data. The first method is a mixture model for analyzing asymmetric mixed-type data. The second is modelling contaminated mixed-type data and identifying potential outliers. Lastly, model averaging techniques are developed for skewed-data based on Occam’s window and parsimonious mixture models. The expectation-maximization algorithm is used here to estimate the model parameters. Both real and simulated data are used for illustration.