<XML><RECORDS><RECORD><REFERENCE_TYPE>3</REFERENCE_TYPE><REFNUM>8985</REFNUM><AUTHORS><AUTHOR>Damoulas,T.</AUTHOR><AUTHOR>Ying,Y.</AUTHOR><AUTHOR>Girolami,M.</AUTHOR><AUTHOR>Campbell,C.</AUTHOR></AUTHORS><YEAR>2008</YEAR><TITLE>Inferring Sparse Kernel Combinations and Relevance Vectors: An application to subcellular localization of proteins</TITLE><PLACE_PUBLISHED>International Conference on Machine Learning and Applications,</PLACE_PUBLISHED><PUBLISHER>IEEE</PUBLISHER><LABEL>Damoulas:2008:8985</LABEL><KEYWORDS><KEYWORD>RVMs</KEYWORD></KEYWORDS<ABSTRACT>In this paper, we introduce two formulations for multi-class multi-kernel relevance vector machines (m-RVMs) that explicitly lead to sparse, both sample-wise and kernel-wise, classification solutions and enable their application to large-scale multi-feature multinomial classification problems where there is an abundance of training samples, classes and feature spaces. The proposed methods are based on an expectation-maximization (EM) framework employing the multinomial probit likelihood and explicit pruning of non-relevant training samples. The resulting relevant vectors are examined in a low-dimensional artificial data-set and we demonstrate the accuracy and sparsity of the method when applied in the challenging bioinformatics problem of predicting protein subcellular localization.</ABSTRACT></RECORD></RECORDS></XML>