Title: Extensions of Non-Negative Matrix Factorization (NMF) to Higher Order Data

Speaker: Morten Morup, Technical University of Denmark

Location: Building 980, Room 95 (Sandia NM)

Brief Abstract: Higher order matrix (tensor) decompositions are mainly used in psychometrics, chemometrics, image analysis, graph analysis and signal processing. For higher order data the two most commonly used decompositions are the PARAFAC and the TUCKER model. If the data analyzed is non-negative it may be relevant to consider additive non-negative components. We here extend non-negative matrix factorization (NMF) to form algorithms for non-negative TUCKER and PARAFAC decompositions. Furthermore, we extend the PARAFAC model to account for shift and echo effects in the data. To improve uniqueness of the decompositions we use updates that can impose sparseness in any combination of modalities. The algorithms developed are demonstrated on a range of datasets spanning from electroencephalography to sound and chemometry signals.

CSRI POC: Tammy Kolda, (925) 294-4679