Low-rank feature selection for multi-view regression

Real life data and information often have different ways to obtain. For example, in computer vision, we can describe an objective by different types, such as text, video and picture. And even from variety of angles. These different descriptors of the same object are usually called multi-view data. In ordinarily, dimensional reduction methods usually include feature selection and subspace learning, respectively, can have better interpretative capability and stabilizing performance, and now are very prevalent method for high-dimensional data. However, it is usually not considering the relationship among class indicators, so the performance of regression model is not very ideal. In this paper, we simultaneously consider feature selection, low-rank selection, and subspace learning into a unified framework. Specifically, under the framework of linear regression model, we first use the low-rank constraint to feature selection which considers two aspects of information inherent in data. The low-rank constraint takes the correlation of response variables into account, then embed an ℓ2, p-norm regularizer to consider the correlation among variety of class indicators, and feature vectors and their corresponding response variables. Meanwhile, we take LDA algorithm which belong to the subspace learning to further adjust relevant feature selection results into account. Lastly, we conducted experiments on several real multi-views image sets and corresponding experimental consequences also validated the furnished method outperformed all comparison algorithms.