Clustering by CP-QR

This paper concerns a novel method to deal with data clustering, which is called completely positive factorizations combined with QR decomposition. We first introduce the concept of completely positive matrices and then the (0,1) CP factorizations, by which some recent work on the linear clustering and the number of the clusters in this case is initialized. A detailed CP-QR algorithm is presented, and its advantage over NMF method is illustrated by an example.