Investigating NMF Initializations: A survey and a microarraycase study.

Nonnegative Matrix Factorization (NMF) revealed to be useful for analysing microarray data and extracting valuable information on complex interactions a– mong biological processes. Due to the iterative nature of most NMF algorithms, initialization is a key aspect as it influences both convergence and final solution. Unlike most studies on NMF initializations, only results on NMF performance in terms of cost function residual are usually presented in literature. And investigations on the influence of NMF initialization techniques on the final solution quality have been conducted only on textual or image datasets. In this paper, we first review many of the initialization schemes from NMF literature and then we present some results on actual biological dataset. We focus on a preliminary enquire on the initialization infuence on latent biological information extraction, when microarray data are mined by NMF. The qualitative analysis performed for the illustrated case study highlighted that NMF extracts an information kernel strictly related to the analysed dataset (i.e. some genes) regardless to the peculiar initialization scheme used to feed it.