Estimation based on RBM from label proportions in large group case

Abstract

Learning a classifier when only knowing about the features and marginal distribution of class labels in each of the data groups is both theoretically interesting and practically useful. Specifically, we are interested in the case where the ratio of the number of data instances to the number of classes is large. For this problem, we show that the performance of a previously proposed discriminative classifier will deteriorate quickly as the ratio grows. In contrast, we formulate a density estimation framework to learn a generative classifier by RBM in this scenario with guaranteed performance under mild assumption. © Springer-Verlag 2013.