Robust deep subspace clustering based on unsupervised fusion learning

Abstract

Subspace clustering methods have been widely used in high-dimensional data analysis due to their excellent capability in processing high-dimensional data. However, traditional methods struggle with nonlinear data, whereas kernel mapping-based methods are limited by kernel function design. Although deep subspace clustering solves this problem to a certain extent, the graph connections generated by its affinity matrix are often over-complete or under-complete. To overcome these limitations, we propose: (1) a flexible encoder based on low-rank embedding that adaptively captures the most informative components in the feature space of the input data while filtering out redundant or noisy information, effectively alleviating over-connected affinities; (2) a dual block-diagonal affinity-inducing self-expression layer that couples a block-diagonal prior with data-driven self-expression, enhancing intra-cluster cohesion, suppressing cross-cluster links and improving structural alignment to ultimately mitigate under-connected structures; and (3) an end-to-end unsupervised clustering architecture, Robust deep subspace clustering based on unsupervised fusion learning (RDSCUF), which fuses the above modules to reinforce interactions between the latent feature space and the structural relationships encoded in the affinity matrix, guiding the network toward more accurate and balanced graph connectivity. The effectiveness and robustness of our method were validated on several widely used real-world datasets.