TY  - CONF
AU  - Muzzarelli, Laura
AU  - Weis, Susanne
AU  - Eickhoff, Simon B.
AU  - Patil, Kaustubh R.
TI  - Rank Selection in Non-negative Matrix Factorization: systematic comparison and a new MAD metric
PB  - IEEE
M1  - FZJ-2019-05146
SP  - 8
PY  - 2019
N1  - This study was partly supported by the Helmholtz Portfolio Theme"Supercomputing and Modeling for the Human Brain" and the EuropeanUnion’s Horizon 2020 Research and Innovation Programme under GrantAgreement No. 785907 (HBP SGA2).
AB  - Non-Negative Matrix Factorization (NMF) is apowerful dimensionality reduction and factorization method thatprovides a part-based representation of the data. In the absence ofa priori knowledge about the latent dimensionality of the data, itis necessary to select a rank of the reduced representation. Severalrank selection methods have been proposed, but no consensusexists on when a method is suitable to use. In this work, we proposea new metric for rank selection based on imputation crossvalidation,and we systematically compare it against six othermetrics while assessing the effects of data properties. Usingsynthetic datasets with different properties, our work criticallyevidences that most methods fail to identify the true rank. Weshow that properties of the data heavily impact the ability ofdifferent methods. Imputation-based metrics, including our newMADimput, provided the best accuracy irrespective of the datatype, but no solution worked perfectly in all circumstances. Oneshould therefore carefully assess characteristics of their dataset inorder to identify the most suitable metric for rank selection.
T2  - 2019 International Joint Conference on Neural Networks (IJCNN)
CY  - 14 Jul 2019 - 19 Jul 2019, Budapest (Hungary)
Y2  - 14 Jul 2019 - 19 Jul 2019
M2  - Budapest, Hungary
LB  - PUB:(DE-HGF)8
DO  - DOI:10.1109/IJCNN.2019.8852146
UR  - https://juser.fz-juelich.de/record/865858
ER  -