Skip to Main content Skip to Navigation

Contributions to fast matrix and tensor decompositions

Abstract : Large volumes of data are being generated at any given time, especially from transactional databases, multimedia content, social media, and applications of sensor networks. When the size of datasets is beyond the ability of typical database software tools to capture, store, manage, and analyze, we face the phenomenon of big data for which new and smarter data analytic tools are required. Big data provides opportunities for new form of data analytics, resulting in substantial productivity. In this thesis, we will explore fast matrix and tensor decompositions as computational tools to process and analyze multidimensional massive-data. We first aim to study fast subspace estimation, a specific technique used in matrix decomposition. Traditional subspace estimation yields high performance but suffers from processing large-scale data. We thus propose distributed/parallel subspace estimation following a divide-and-conquer approach in both batch and adaptive settings. Based on this technique, we further consider its important variants such as principal component analysis, minor and principal subspace tracking and principal eigenvector tracking. We demonstrate the potential of our proposed algorithms by solving the challenging radio frequency interference (RFI) mitigation problem in radio astronomy. In the second part, we concentrate on fast tensor decomposition, a natural extension of the matrix one. We generalize the results for the matrix case to make PARAFAC tensor decomposition parallelizable in batch setting. Then we adapt all-at-once optimization approach to consider sparse non-negative PARAFAC and Tucker decomposition with unknown tensor rank. Finally, we propose two PARAFAC decomposition algorithms for a classof third-order tensors that have one dimension growing linearly with time. The proposed algorithms have linear complexity, good convergence rate and good estimation accuracy. The results in a standard setting show that the performance of our proposed algorithms is comparable or even superior to the state-of-the-art algorithms. We also introduce an adaptive nonnegative PARAFAC problem and refine the solution of adaptive PARAFAC to tackle it. The main contributions of this thesis, as new tools to allow fast handling large-scale multidimensional data, thus bring a step forward real-time applications.
Document type :
Complete list of metadatas
Contributor : Abes Star :  Contact
Submitted on : Tuesday, February 20, 2018 - 10:59:37 AM
Last modification on : Wednesday, November 20, 2019 - 1:42:44 AM
Long-term archiving on: : Monday, May 21, 2018 - 12:16:27 PM


Version validated by the jury (STAR)


  • HAL Id : tel-01713104, version 1


Viet-Dung Nguyen. Contributions to fast matrix and tensor decompositions. Other. Université d'Orléans, 2016. English. ⟨NNT : 2016ORLE2085⟩. ⟨tel-01713104⟩



Record views


Files downloads