Divergences for prototype-based classification and causal structure discovery: Theory and application to natural datasets

This thesis is a two-part thesis. In the first part we discuss extensions of LVQ prototype-based classifiers that use information theoretic measures as distance measures. We also present work on the use of different representations of data, in this case histograms of images, SIFT and SURF features in the same LVQ system. We show the possibility of formulating one combined distance measure for the heterogeneous dataset formed by a combination of their individual representative distance measures.

In the second part we delve into causal structure discovery and its application to real world problems. We also presented a first attempt at leveraging some of the techniques of causal learning and applying them to feature relevance learning in LVQ. We also present some deployment examples of some of the techniques to solve real world problems in Uganda particularly using causal analysis to predict the state of food security at the household level and automated diagnosis of crop disease using divergence-based LVQ techniques on a mobile device.