Affiliations: [a] CPIB, School of Bioscience, The University of Nottingham, Nottingham, UK | [b] School of Computer Science, The University of Nottingham, Nottingham, UK
Correspondence:
[*]
Corresponding author: Jianyong Sun, CPIB, School of Bioscience, The University of Nottingham, Nottingham, UK. Tel.: +44 (0)11 5951 6108; E-mail: jsun@cpib.ac.uk.
Abstract: Both parametric Bayesian mixture and non-parametric Dirichlet process mixture modelling (DPM) approaches for density estimation and clustering allow for automatic model selection. It is interesting to study which approach can better fit the data. In this paper, we focus on robust clustering taking the Student t-distribution as the building block. We develop two novel robust clustering algorithms, one using Type-IV Student t-distribution mixture modelling (SMM) and one robust DPM (RDPM), and explain them in detail. The new algorithms are compared using controlled experiment settings and benchmark UCI datasets, in terms of commonly-used internal and external cluster validity indices. Experimental results show that Type-IV SMM shows comparable performance to Type-II SMM, while additionally identifying outliers, and that RDPM outperforms conventional DPM. When comparing the two new algorithms with each other, they are found to perform comparably, but Type-IV SMM is less sensitive to initialisation and has a better generalisation ability. Hence, it is recommended to use Type-IV SMM for robust clustering and model selection.
