Robust Multi-label Feature Selection via Low-dimensional Embedding and Manifold Regularization

Abstract

Data dimension is one of the important factors that affect the efficiency of multi-label learning. In order to reduce the dimension of multilabel data, feature selection technology has been widely used in multi-label learning to obtain the most discriminant feature subset. In this paper, a robust multilabel feature selection based on low-dimensional embedding and manifold learning is proposed and named REMFS. Firstly, a low-dimensional embedding is constructed to mine the relationship between feature space and label space. And then we make the low-dimensional embedding and ground-truth label distribution as consistent as possible. In addition, manifold learning is introduced to preserve local geometry for features and labels simultaneously. Then, an iterative optimization scheme is adopted to solve the objective function of REMFS. Finally, comparison experiments are carried out on multiple datasets to prove the effectiveness of REMFS method.