Deep subspace clustering: A mechanism-oriented survey

High-dimensional, sparse, and nonlinear data pose substantial challenges to conventional clustering methods. Deep subspace clustering (DSC) addresses these challenges by integrating nonlinear representation learning with subspace structure modeling. This survey presents a unified, mechanism-oriented review of DSC and proposes a two-dimensional taxonomy based on view setting (single-view or multi-view) and latent-space property (linear or nonlinear). Representative methods are systematically examined in terms of representation learning, self-expression and affinity construction, structural regularization, cross-view collaboration, and optimization. We further identify key challenges involving complex data distributions, incomplete and heterogeneous views, noise robustness, and computational scalability. Accordingly, we outline future directions in scalable self-expression, robust structural learning under imperfect data, semantic and theoretical analysis, and standardized benchmarking. This survey provides a systematic framework for understanding the methodological evolution and design trade-offs of DSC.