Adaptive structure learning for semi-supervised feature selection with binary single-label learning

Abstract

Learning pseudo-labels for unlabeled samples provides more helpful information in semi-supervised feature selection (SSFS), and the labels of unlabeled samples are learned as continuous values by most existing SSFS methods. Whereas the given labels of labeled samples are encoded in a one-hot encoding way, the two are not uniform in form and do not provide more explicit supervised information. So, this paper introduces binary single-label learning, which learns unlabeled sample labels into a uniform one-hot encoding form. Furthermore, this paper preserves the data's local and global structure by combining improved Euclidean distance-based adaptive graph learning with sparse representation learning. A novel SSFS model called Adaptive Structure Learning for Semi-supervised Feature Selection with Binary Single-label Learning (ASBLFS) is proposed, and an efficient optimization algorithm is derived. Finally, the following conclusions are observed through extensive experiments with several advanced SSFS models on 15 benchmark datasets: (1) Binary single labels achieve better performance than continuous labels on some datasets, suggesting that binary labels can provide more explicit supervisory information. (2) ASBLFS shows the second-best or best performance on most datasets, demonstrating the superiority of ASBLFS.