Abstract:

To overcome the deficiency of Laplacian score algorithm which makes feature selection mostly depending on the local geometrical structure of samples, an improved semi-supervised feature selection algorithm was proposed, based on the constraint Laplacian score. It utilized the cannot-link pairwise constraints among samples as the global structure. Then the selected features were those can preserve the local structure from the nearest neighbor graph,and preserve the global structure from the cannot-link constraints. Experiments on Yale and PIE(Fave pose,Illamination,Expression dadbase) datasets show that the performance of proposed algorithm outperformed Laplacian score algorithm significantly, and was equivalent to the supervised Fisher score algorithm and the latest semi-supervised constraint score algorithms. And it is even better than constraint score algorithm in terms of stability.

Key words: feature selection, local structure information, cannot-link constraint, semi-supervised learning