Semi-Supervised Discriminant Analysis Based On Data Structure

Dimensionality reduction is a key data-analytic technique for mining high-dimensional data. In this paper, we consider a general problem of learning from pairwise constraints in the form of must-link and cannotlink. As one kind of side information, the must-link constraints imply that a pair of instances belongs to the same class, while the cannot-link constraints compel them to be different classes. Given must-link and cannot-link information, the goal of this paper is learn a smooth and discriminative subspace. Specifically, in order to generate such a subspace, we use pairwise constraints to present an optimization problem, in which a least squares formulation that integrates both global and local structures is considered as a regularization term for dimensionality reduction. Experimental results on benchmark data sets show the effectiveness of the proposed algorithm.