Consistency of anchor-based spectral clustering

Abstract Anchor-based techniques reduce the computational complexity of spectral clustering algorithms. Although empirical tests have shown promising results, there is currently a lack theoretical support for anchoring approach. We define specific anchor-based algorithm and show that it amenable to rigorous analysis, as well being effective in practice. establish consistency method an asymptotic setting where data sampled from underlying continuous probability distribution. In particular, we provide sharp conditions number nearest neighbors algorithm, which ensure can recover with high disjoint clusters are mutually separated by positive distance. illustrate performance on synthetic explain how convergence analysis be used inform practical choice parameter scalings. also test accuracy efficiency two large scale real sets. find offers clear advantages over standard clustering. competitive state-of-the-art LSC Chen Cai (Twenty-Fifth AAAI Conference Artificial Intelligence, 2011), while having added benefit guarantee.