In response to the increasing availability of high-dimensional data, effectively uncovering the underlying low-dimensional structures has become a critical yet challenging task. We introduce a unified optimization framework, termed SVSS, that simultaneously learns robust data representations and an adaptive similarity graph for single-view subspace segmentation. By explicitly decomposing the raw data into clean and noisy components, our method constructs a more reliable affinity matrix without manual tuning. Furthermore, we extend this framework to the multi-view scenario, namely MVSS, where a consensus similarity matrix shared across all views is jointly inferred. Experimental results on various benchmark datasets demonstrate that both SVSS and MVSS outperform state-of-the-art alternatives in terms of clustering accuracy and normalized mutual information.