Classification of partially labelled data using "mixture of expert" models

Thesis

In this thesis, we are concerned with the classification of partially labeled data. By partially labeled data we mean data where measurements are available from experimental units which are known to belong to one of a set of known classes but whose individual membership to subclasses within the known class is not known. Examples of these applications include fisheries research where fish lengths are available but sexual identities are not, sedimentology where information is available about the grain size distribution of a sample of sand but not its mineral composition and medical diagnosis

where the symptoms of the patients are known but not disease classifications.

A popular way of handling such partially labelled data is to use a mixture of Gaussian densities. Relying on the assumption of Gaussian densities (more information), results in an more efficient (less variance) discrimination procedure if the modelling assumptions are satisfied. However in practice, these assumptions rarely hold and often some of the features are qualitative variables, and hence it is generally of the view that logistic discrimination is a more robust bet as it relies on fewer assumptions. We chose to use a mixture of logistic regressions, embedded within a hierarchical mixture of experts to classify our data.

We compared its use in the plug-in-approach (frequentist approach) versus the predictive approach (Bayesian approach) in classification. The density of parameters required for the predictive approach was obtained using Markov Chain Monte Carlo simulation.

Authors: Teo, T

• DOI: 10.5287/ora-bxz5bnnxd
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford