Unsupervised Spectral Feature Selection with Dynamic Hyper-graph Learning

Abstract

Unsupervised Spectral Feature Selection (USFS) methods may output interpretable and discriminative results by integrating a Laplacian regularizer into the feature selection framework to keep the similarity of the training samples. To do this, USFS methods usually construct the Laplacian matrix using either a general graph or a Hyper-graph on the original data. Since the construction of the Laplacian matrix is separated from the process of feature selection and the original data usually contain noise, the performance of feature selection will be influenced. Based on previous observation that a hyper-graph measuring the relationship among more than two samples is able to characterize more complex (i.e. high-order) relationship of the samples than a general graph measuring the relationship between two samples, in this paper, we propose a novel feature selection method by dynamically learning a hyper-graph based Laplacian matrix to measure the relationship among the samples for selecting the features. Experimental results on real datasets showed that our proposed method outperforms the state-of-the-art feature selection methods in terms of both clustering and segmentation performance.