Structured graph learning for clustering and semi-supervised classification

• A graph learning framework, which captures both the global and local structure in data, is proposed. Theoretical analysis builds connections of our model to k-means, spectral clustering, kernel k-means. Extensions semi-supervised classification multiple are presented. Graphs have become increasingly popular modeling structures interactions a wide variety problems during last decade. Graph-based clustering techniques shown impressive performance. This paper proposes framework preserve data. Specifically, method uses self-expressiveness samples capture adaptive neighbor approach respect structure. Furthermore, most existing graph-based methods conduct on learned from original data matrix, doesn’t explicit cluster structure, thus they might not achieve optimal By considering rank constraint, achieved will exactly c connected components if there clusters or classes. As byproduct this, label inference jointly iteratively implemented principled way. Theoretically, we show that equivalent combination k -means under certain condition. Extensive experiments demonstrate proposed outperforms other state-of-the-art methods.