Dimensionality reduction algorithm of tensor data based on orthogonal tucker decomposition and local discrimination difference

Dimensionality Reduction (DR) is a significant subject which have aroused extensive attention of researchers. In this paper, a novel method is proposed to reduce the dimensionality of tensor data based on orthogonal Tucker decomposition model and local discrimination difference regularization (OTDLDD-TDR). The proposed method not only defines the relation between high-dimensional and low-dimensional data spaces from the global view point, but also well preserves both geometric distribution (internal attribute) and tag information (external attribute) belonging to data from the local perspective. For the purpose of modeling the combination of these two types of information, we decompose the tensor data into individual parts on the basis of data distribution in sample space. Then, we calculate the local discrimination difference with regard to each part via the categories assigned in specific application scenarios. We seek to minimize the output of local discrimination difference, and impose it as a regularization constraint on the orthogonal Tucker decomposition model to finally achieve DR for tensor. The proposed method is the first attempt to organically combine tensor Tucker decomposition and local discrimination difference. We conducted comparison experiments for the proposed method, state-of-the-art methods published in recent years, and classic and representative methods. The experimental results in data classification and clustering on eight real-world datasets demonstrate the effectiveness of OTDLDD-TDR.