Supervised Generative Models for Learning Dissimilarity Data

Exemplar based techniques such as affinity propagation [1] represent data in terms of typical exemplars. This has two benefits: (i) the resulting models are directly interpretable by humans since representative exemplars can be inspected in the same way as data points, (ii) the model can be applied to any dissimilarity measure including non-Euclidean or non-metric settings. Most exemplar based techniques have been proposed in the unsupervised setting only, such that their performance in supervised learning tasks can be weak depending on the given data. Here, we address the problem of learning exemplar-based models for general dissimilarity data in a discriminative framework. For this purpose, we extend a generative model proposed in [2] to an exemplar based scenario using a generalized EM framework for its optimization. The resulting classifiers represent data in terms of sparse models while keeping high performance in state-of-the art benchmarks.