Multifactor sparse feature extraction using Convolutive Nonnegative Tucker Decomposition

Qiang Wu, Liqing Zhang, Andrzej Cichocki

Research output: Contribution to journalArticlepeer-review

13 Citations (Scopus)


Multilinear algebra of the higher-order tensor has been proposed as a potential mathematical framework for machine learning to investigate the relationships among multiple factors underlying the observations. One popular model Nonnegative Tucker Decomposition (NTD) allows us to explore the interactions of different factors with nonnegative constraints. In order to reduce degeneracy problem of tensor decomposition caused by component delays, convolutive tensor decomposition model is an appropriate model for exploring temporal correlations. In this paper, a flexible two stage algorithm for K-mode Convolutive Nonnegative Tucker Decomposition (K-CNTD) model is proposed using an alternating least square procedure. This model can be seen as a convolutive extension of Nonnegative Tucker Decomposition. The patterns across columns in convolutive tensor model are investigated to represent audio and image considering multiple factors. We employ the K-CNTD algorithm to extract the shift-invariant sparse features in different subspaces for robust speaker recognition and Alzheimer's Disease(AD) diagnosis task. The experimental results confirm the validity of our proposed algorithm and indicate that it is able to improve the speaker recognition performance especially in noisy conditions and has potential application on AD diagnosis.

Original languageEnglish
Pages (from-to)17-24
Number of pages8
Publication statusPublished - 10 Apr 2014
Externally publishedYes


  • Alternating Least Squares (ALS)
  • Convolutive Tucker Model
  • Feature extraction
  • Nonnegative Tensor Decomposition
  • Robustness


Dive into the research topics of 'Multifactor sparse feature extraction using Convolutive Nonnegative Tucker Decomposition'. Together they form a unique fingerprint.

Cite this