Abstract: Traditional subspace clustering algorithms express the similarity between data by learning the self-expression matrix. However, this strategy fails to deal with large-scale datasets and out-of-sample problems. A deep embedding subspace clustering network is proposed to address these limitations in this paper. Firstly, the autoencoder is employed to obtain the latent representation of the original data. Then, the similarity between latent representations is calculated by a predefined function. Thereby, the self-expression matrix is constructed correspondingly. Finally, the spectral clustering algorithm is utilized to obtain the clustering results. The proposed model avoids directly learning the self-expression coefficients between the data and instead chooses to obtain the latent representation and inherent similarity of the data through the mapping function, which has a wider field of application. Experimental results on four publicly available large-scale datasets demonstrate that the proposed model achieves state-of-the-art results in both accuracy and adjusted rand index, with an average ranking of 1.25 in normalized mutual information. Parameter sensitivity and generalization experiments further validate robustness and ability to handle out-of-sample problems.