Modified Constraint Scores for Semi-Supervised Feature Selection

Semi-supervised constraint scores, which utilize both pairwise constraints and the local property of the unlabeled data to select features, achieve comparable performance to the supervised feature selection methods. The local property is characterized without considering the pairwise constraints and these two conditions are introduced independently. However, the pairwise constraints and the local property may contain conflicting information. In this paper, we utilize the conflicting information to improve the local property. Instead of characterizing the local property by all neighbors, samples which do not appear in the cannot-link constraints can be used. A performance indicator, called neighborhood-cannotlink (NC) coefficient, is proposed to measure the improvement of the local property. We use the improved local property and the pairwise constraints to perform semi-supervised constraint scores algorithm. Experiments on several real world data sets demonstrate the effectiveness of the methods.