B0396
Title: On a novel Bayesian nonparametric approach to supervised learning for binary data
Authors: Jose Antonio Perusquia Cortes - University of Kent (United Kingdom) [presenting]
Jim Griffin - University College London (United Kingdom)
Cristiano Villa - Newcastle University (United Kingdom)
Abstract: Supervised learning models provide a powerful tool for the classification of unlabeled observations. However, most of the classifiers have been built on a discriminative approach. Hence, they cannot provide an understanding of the generative process of the data. That is why the rich probabilistic background of Bayesian nonparametric models yield an interesting approach to supervised and unsupervised classification. We centre our attention on exploring a novel methodology to supervised classification for binary data using a beta compound random measure as a building block. This Bayesian nonparametric prior allows us to fully characterise the distribution of the different groups through the means of a score distribution that modifies the jumps of a directing beta process and hence, identify not only the most influential features overall but for each group as well.
