Non-negative sub-tensor ensemble factorization (NsTEF) algorithm. A new incremental tensor factorization for large data sets.

Vincent Vigneron (1) , Andreas Kodewitz (1) , Michele Nazareth da Costa (2) , Ana Maria Tome (3) , Elmar Langlang (4)

1 IBISC - Informatique, Biologie Intégrative et Systèmes Complexes
2 DSPCom - Digital Signal Processing and Communications laboratory [Campinas]
3 IEETA - Institute of Electronics and Telematics Engineering of Aveiro
4 Institut für Biophysik und physikalische Biochemie

In this work we present a novel algorithm for nonnegative tensor factorization (NTF). Standard NTF algorithms are very restricted in the size of tensors that can be decomposed. Our algorithm overcomes this size restriction by interpreting the tensor as a set of sub-tensors and by proceeding the decomposition of sub-tensor by sub-tensor. This approach requires only one sub-tensor at once to be available in memory.

Mots clés: CANDECOMP/PARAFAC Decomposition; Incremental algorithm; Learning method; Matrix factorization; Non-negative tensor decomposition; NTF

Domaines: Apprentissage [cs.LG]; Traitement du signal et de l'image [eess.SP]

Vincent Vigneron, Andreas Kodewitz, Michele Nazareth da Costa, Ana Maria Tome, Elmar Langlang. Non-negative sub-tensor ensemble factorization (NsTEF) algorithm. A new incremental tensor factorization for large data sets.. Signal Processing, 2018, 144, pp.77-86. ⟨10.1016/j.sigpro.2017.09.012⟩. ⟨hal-01629626⟩