TENSOR-BASED ALGEBRA FOR MULTILINEAR STRUCTURE OF MICROARRAY EXPRESSION RECOGNITION

Abstract

In this paper, there are two current challenges of microarray expression are investigated, namely missing values recovery and feature extraction for supervised learning. The first axis is focus on how to deal with the main properties of time sequences of microarray, including tensor structure, noise, and temporal dynamic characteristics. This allows discovering latent factors and evolving trends for offering missing imputation. Then, an improvement of orthogonal Tucker decomposition based on discriminant analysis for multilinear structure of microarray expression is presented for dimensionality reduction. Consequently, an exploring a novel type of third-order microarray expression, termed as gene - sample - time (GST), is presented for biological sample classification. The contributions will be distributed along two main thrusts of effectiveness; including latent modeling setting for imputing missing values based on the High-Order Kalman Filter and feature extraction based on Tensor Discriminative Feature Extraction. Those proposal methods are carried out on Interferon beta (INFβ) dataset of GST-microarray expressions to distinguish the patients in the favorable response group and the remaining patients in the problematic-treatment-response group. The experimental performance corroborates the advantages of the proposed approaches upon those of the matrix-based algorithms and recent tensor-based, discriminant-decomposition, in terms of missing values completion, classification accuracy of 90.23% and computation time.