Unsupervised Feature Selection With Ordinal Preserving Self-Representation

Unsupervised feature selection is designed to select an optimal feature subset without any label information from high-dimensional data, which is implemented by eliminating the irrelevant and redundant features and has been attracted widespread attention in recent years. Specifically, the obtained low-dimensional representation is interpretable that is useful to machine learning applications. In this paper, we propose a novel unsupervised feature selection algorithm, namely ordinal preserving self-representation (OPSR) for image classification and clustering. First, each feature in high-dimensional data is represented by the linear combination of other features. Then, the topology information is introduced into the objective function for utilizing the ordinal locality of high-dimensional data adequately. At last, an efficient iteratively update algorithm is designed to solve the proposed OPSR, and its convergence is proved in detail. Extensive experimental results on six benchmark databases demonstrate that the effectiveness of the OPSR and its superiority also is verified by comparing with some state-of-the-art feature selection algorithms.