Multilayer concept factorization for data representation

Abstract

Previous studies have demonstrated that Concept Factorization (CF) have yielded impressive results for dimensionality reduction and data representation. However, it is difficult to get a desired result by using single layer concept factorization for some complex data, especially for ill-conditioned and badly scaled data. To improve the performance of the existing CF algorithms, in this paper, we proposed a novel clustering approach, called Multilayer Concept Factorization (MCF), for data representation. MCF is a cascade connection of L mixing subsystems to decompose the observation matrix iteratively in a number of layers. Meanwhile, we explore the corresponding update solutions of the MCF method to reduce the risk of getting stuck in local minima in non-convex alternating computations. Experimental results on document and face dataset demonstrate that our proposed method achieves better clustering performance in terms of accuracy and normalized mutual information in clustering.