From Global to Local: Semantic-Aware Instance-Wise Feature Selection

Feature selection is a promising dimension reduction technology that focuses on a reduced subspace by selecting excellent features. Most existing approaches tend to emphasize the discriminative ability of features based on either a global or a local evaluation criterion alone, and a few holistic approaches explore their selection granularity beyond the instance level. This study presents a novel Semantic-aware Instance-wise Feature selection model, dubbed SIF, to address the weakness of existing methods, which assess the significance of features from an individual view. Furthermore, SIF proposes to specify feature representations at the instance level, which is rarely touched by existing methods given the considerable learning complexity. In particular, SIF is designed as a sequential pipeline framework. First, it explicitly models semantic correlations and employs this information to select semantic-aware features. Then, inconsistent instances are captured and guide the instance-wise feature selection. Both types of features constitute the final optimal feature subset, which can represent semantics at a global level as well as describe instance characteristics at a local level. An extensive experimental evaluation illustrates the superiority of SIF under various metrics.