Research on exponential regularization approach for nonnegative Tucker3 decomposition

Methods of nonnegative tensor factorization (NTF), such as NTF1, NTF2, etc., are extension of nonnegative matrix factorization (NMF) for multi-way data analysis. As an existing NTF method, nonnegative Tucker3 decomposition (NTD) is researched for three-way decomposition in this paper. Firstly, an approach utilizing matrix exponentials built on Tikhonov-type regularization to enforce sparseness is proposed to extract image features instead of exclusively using Tucker tensor decomposition. Meanwhile, updating algorithms, derived from updating rules of NMF, are allowed to efficiently implement updating of mode matrices and core tensors alternatively for accuracy. Then, experimental cases of alternating least squares (ALS) and conjugate nonnegative constraints, called nonnegative alternating least squares (NALS), are studied to remedy data overfitting in computing procedures. Lastly, the proposed method exhibits more advantageous results than other algorithms of Tucker3 for feature extraction, thanks to computer simulations performed in the context of data analysis.