Robust Subspace Learning with Double Graph Embedding

Low-rank-based methods are frequently employed for dimensionality reduction and feature extraction in machine learning. To capture local structures, these methods often incorporate graph embedding, which requires constructing a zero-one weighted neighborhood graph to extract local information from the original data. However, these methods are incapable of learning an adaptive graph that reveals intricate relationships among distinct samples within noisy data. To address this issue, we propose a novel unsupervised feature extraction method called Robust Subspace Learning with Double Graph Embedding (RSL_DGE). RSL_DGE incorporates a low-rank graph into the graph embedding process to preserve more discriminative information and remove noise simultaneously. Additionally, the l(2,1)-norm constraint is also imposed on the projection matrix, making RSL_DGE more flexible in selecting feature dimensions. Several experiments demonstrate that RSL_DGE achieves competitive performance compared to other state-of-the-art methods.