Multilinear local discriminant analysis using adaptive neighborhood graph construction

In this paper we introduce a novel supervised dimensionality reduction technique called multilinear local discriminant analysis, which can preserve the local geometrical and discriminant structure of data with tensor representation. Firstly, we adaptively choose the neighbors of the sample and construct within-class and between-class neighborhood graph based on sample density and similarity. Then, define the local within-class and between-class scatter matrix measured in tensor metric. Ultimately iteratively gain optimal subspace by k-mode optimization, which maximize the local within-class scatter and at the same time minimize the between-class scatter by unfolding the tensor along different tensor direction. Experimental results on ORL face database validate the effectiveness of the proposed method.