Sparse and low-redundant subspace learning-based dual-graph regularized robust feature selection

摘要

Feature selection can reduce the dimension of data and select the representative features. The available researches have shown that the underlying geometric structures of both the data and the feature manifolds are important for feature selection. However, few feature selection methods utilize the two geometric structures simultaneously in subspace learning. To solve this issue, this paper proposes a novel algorithm, called sparse and low-redundant subspace learning-based dual-graph regularized robust feature selection (SLSDR). Based on the framework of subspace learning-based graph regularized feature selection, SLSDR extends it by introducing the data graph. Specifically, both data graph and feature graph are introduced into subspace learning, so SLSDR preserves the geometric structures of the data and feature manifolds, simultaneously. Consequently, the features which best preserve the manifold structures are selected. Additionally, the inner product regularization term, which guarantees the sparsity of rows and considers the correlations between features, is imposed on the feature selection matrix to select the representative and low-redundant features. Meanwhile, the -norm is imposed on the residual matrix of subspace learning to ensure the robustness to outlier samples. Experimental results on twelve benchmark datasets show that the proposed SLSDR is superior to the six state-of-the-art algorithms from the literature.