Constrained feature weighting for semi-supervised learning

Semi-supervised feature selection plays a crucial role in semi-supervised classification tasks by identifying the most informative and relevant features while discarding irrelevant or redundant features. Many semi-supervised feature selection approaches take advantage of pairwise constraints. However, these methods either encounter obstacles when attempting to automatically determine the appropriate number of features or cannot make full use of the given pairwise constraints. Thus, we propose a constrained feature weighting (CFW) approach for semi-supervised feature selection. CFW has two goals: maximizing the modified hypothesis margin related to cannot-link constraints and minimizing the must-link preserving regularization related to must-link constraints. The former makes the selected features strongly discriminative, and the latter makes similar samples with selected features more similar in the weighted feature space. In addition, L1-norm regularization is incorporated in the objective function of CFW to automatically determine the number of features. Extensive experiments are conducted on real-world datasets, and experimental results demonstrate the superior effectiveness of CFW compared to that of the existing popular supervised and semi-supervised feature selection methods.