Semi-supervised Regularization Learning
Enliang Hu · Journal of Chinese Computer Systems · 2010
In this paper,we study semi-supervised linear dimensionality reduction algorithm.Beyond conventional methods which merely consider labeled instances,the semi-supervise scheme allows to both the side information and abundant unlabeled instances into learning so as to achieve better generalization performance.Under semi-supervised setting,our objective is to learn a smooth and discriminative subspace.Specifically,cannot-link pairwise constraints are used to maximize the distance between instances from different classes,while must-link pairwise constraints are used to minimize the distance between instances from the same class.Additionally,we consider both the geometrical structure of the data and feature structure of projecting vector as the regularization terms to guide the process of dimensionality reduction.Moreover,the proposed algorithm can easily solve out-of-sample problem.Experimental results demonstrate the efficiency and effectiveness of our algorithm.