Abstract: The deep subspace clustering algorithm utilizes deep neural networks to map the original input data to a latent space and employs the self-expression of the data as a measure of data similarity, effectively achieving clustering of high-dimensional data. However, such algorithms only focus on the self-expressive relationship in the latent space, resulting in their performance heavily relying on the quality of features extracted by the deep neural networks. Additionally, the regularization process ignores the connectivity within each subspace, affecting the performance of spectral clustering. To address these issues, a deep subspace clustering algorithm based on dual self-expression and the maximum entropy principle is proposed. This algorithm simultaneously learns the self-expressive relationships in both the latent space and the input space, guiding the deep neural network to obtain data representations suitable for subspace clustering. By maximizing the entropy of the similarity matrix, it ensures that elements within the same subspace are uniformly and densely distributed, thereby improving the performance of data clustering. Extensive experiments on five datasets verify the effectiveness of the proposed algorithm. 

Key words: subspace clustering, self-expression, deep neural network, the maximum entropy principle