Semi-Supervised Dimensionality Reduction with Pairwise Constraints Using Graph Embedding for Face Analysis

Following the intuition that the image variation of faces can be effectively modeled by low dimensional linear spaces, we propose a novel linear subspace learning method for face analysis in the framework of graph embedding model, called semi-supervised graph embedding (SGE). This algorithm builds an adjacency graph which can best respect the geometry structure inferred from the must-link pairwise constraints, which specify a pair of instances belong to the same class. The projections are obtained by preserving such a graph structure. Using the notion of graph Laplacian, SGE has a closed solution of an eigen-problem of some specific Laplacian matrix and therefore it is quite efficient. Experimental results on Yale standard face database demonstrate the effectiveness of our proposed algorithm.