Clustering by non-negative matrix factorization with independent principal component initialization

Non negative matrix factorization (NMF) is a dimensionality reduction and clustering method, and has been applied to many areas such as bioinformatics, face images classification, and so on. Based on the traditional NMF, researchers recently have put forward several new algorithms on the initialization area to improve its performance. In this paper, we explore the clustering performance of the NMF algorithm, with emphasis on the initialization problem. We propose an initialization method based on independent principal component analysis (IPCA) for NMF. The experiments were carried out on the four real datasets and the results showed that the IPCA-based initialization of NMF gets better clustering of the datasets compared with both random and PCA-based initializations.