An Improved Consensus Clustering for Nonnegative Matrix Factorization in Molecular Cancer Class Discovery

Recently nonnegative matrix factorization (NMF) has been proven powerful for nonnegative data analysis, especially in analyzing gene expression data. We propose an modified consensus clustering mechanism with soft sample assignment to improve the clustering accuracy. The idea is to use normalized inner product or cosine similarity matrix for the connectivity matrix of the consensus clustering. The experimental results demonstrate the effectiveness of the proposed method.