Log-Quad divergence for Non-negative Matrix Factorization in multi-model prediction
Log-Quad divergence for Non-negative Matrix Factorization in multi-model prediction

Summary/Abstract: The aim of this paper is to present a new Non-negative Matrix Factorization (NMF) algorithm based on Log-Quad divergence, and to demonstrate its application to the separation of latent destructive components contained in prediction results in a multi-model approach. We provide an example of its application to a real economic problem, i.e. forecasting electricity consumption on the basis of information about hourly use of electricity in Poland in the period of 1988–1997. We evaluated and compared this method with other blind signal (source) separation techniques, such as Independent Component Analysis (ICA) and Algorithm for Multiple Unknown Signals Extraction (AMUSE). The results show that the NMF algorithm based on Log-Quad divergence has an interesting ability to improve predictions for small volumes of data.

• Details
• Contents