Graph Regularized Projective Non-negative Matrix Factorization for Image Clustering

For enhancing the cluster accuracy, this paper presents a novel algorithm called Graph regularized Projective Non-negative Matrix Factorization (GPNMF). When original data is factorized in lower dimensional space using NMF, GPNMF preserves the local structure and intrinsic geometry of data, using a PCA-like regularization term to ensure the projection dose not lose too much information available in the original domain. An efficient multiplicative updating procedure was produced, the relation with gradient descent method showed that the updating rules are special case of its. Compared with NMF and its improved algorithms, experiment results on USPS handwrite database and ORL face database have shown that the proposed method achieves better clustering results.