Learning a Kernel Matrix Using Some Similar and Dissimilar pairs

Abstract

A lot of machine learning algorithms are based on metric functions, which good functions lead to better results. Distance metric learning has been widely attracted by researchers in last decade. Kernel matrix is somehow a distance function which indicates the similarity between two instances in the feature space which contains high dimensions. Traditional distance metric learning approaches are based on Mahanalobis distance which result in optimizing a positive semi definite problem. This kind of approaches need high computational time and do not work well in the case of data with high dimensions. Another filed which is involved by researchers in last decade is building a good kernel matrix which separate non separable data best. This paper proposed a new algorithm in order to learn kernel matrix which is based on distance metric learning. It is implemented and applied to several standard data sets and the results are shown.