A New Framework for Dissimilarity and Similarity Learning

Abstract

In this work we propose a novel framework for learning a (dis)similarity function. We cast the learning problem as a binary classification task or a regression task in which the new learning instances are the pairwise absolute differences of the original instances. Under the classification approach the class label we assign to a specific pairwise difference indicates whether the two original instances associated with the difference are members of the same class or not. Under the regression approach we assign positive target values to the pairwise differences of instances from different classes and negative target values to the differences of instances of the same class. The computation of the (dis)similarity of two examples amounts to the computation of prediction scores for classification, or the prediction of a continuous value for regression. The proposed framework is very general as we are free to use any learning algorithm. Moreover, our formulation generally leads to a (dis-)similarity which, depending on the learning algorithm, can be efficient and simple to learn. Experiments performed on a number of classification problems demonstrate the effectiveness of the proposed approach.