1: \begin{abstract}

2: 

3:     Anchor-based techniques reduce the computational

4:     complexity of spectral clustering algorithms.

5:     Although empirical tests have shown promising results, there is currently a lack of theoretical support for the anchoring approach.

6:     We define a specific anchor-based algorithm and show that it

7:   is amenable to rigorous analysis, as well as being effective in practice.

8:       We establish the theoretical consistency of the method

9:        in an asymptotic setting where data is sampled

10:         from an underlying continuous probability distribution.

11:          In particular, we provide sharp asymptotic conditions for the algorithm parameters

12:          which ensure that

13:           the anchor-based method can recover with high probability

14:            disjoint clusters that are mutually separated by a positive distance.

15:             We illustrate the

16:             performance of the algorithm on synthetic

17:              data and explain how the

18:               theoretical convergence analysis can be used to inform the

19:              practical choice of parameter scalings.

20:              We also test the accuracy and efficiency of the algorithm on two large scale real data sets.

21:              We find that the algorithm offers clear advantages over standard spectral clustering.

22:               We also find that it

23:              is competitive with the state-of-the-art

24:              LSC method of Chen and Cai (Twenty-Fifth AAAI Conference on Artificial Intelligence, 2011), while having the added

25:              benefit of a consistency guarantee.

26: 

27: \end{abstract}

28: