Unsupervised feature selection with adaptive multiple graph learning

Abstract

Unsupervised feature selection methods try to select features which can well preserve the intrinsic structure of data. To represent such structure, conventional methods construct various graphs from data. In most cases, those different graphs often contain some consensus and complementary information. To make full use of such information, we construct multiple base graphs and learn an adaptive consensus graph from these base graphs for feature selection. In our method, we integrate the multiple graph learning and the feature selection into a unified framework, which can jointly characterize the structure of the data and select the features to preserve such structure. The underlying optimization problem is hard to solve, and we solve it via a block coordinate descent schema, whose convergence is guaranteed. The extensive experiments well demonstrate the effectiveness of our proposed framework.