Clustering via binary embedding

In this paper, we present a novel clustering scheme based on binary embeddings, which provides compact and informative binary representations of high-dimensional objects. The binary representations are obtained with a collection of one-class classifiers learned from (pseudo) randomly selected points in the dataset. To cluster the binary representations, we consider two approaches: a mixture of Bernoulli distributions and a recent biclustering approach called CRAFT. The empirical evaluation in comparison with both classic and recent clustering methods, based on 12 different datasets, provides encouraging results. The main feature of the proposed method is that it is agnostic to the shape of the clusters.