Clustering is the problem of separating a set of objects into groups (calledclusters) so that objects within the same cluster are more similar to eachother than to those in different clusters. Spectral clustering is a nowwell-known method for clustering which utilizes the spectrum of the datasimilarity matrix to perform this separation. Since the method relies onsolving an eigenvector problem, it is computationally expensive for largedatasets. To overcome this constraint, approximation methods have beendeveloped which aim to reduce running time while maintaining accurateclassification. In this article, we summarize and experimentally evaluateseveral approximation methods for spectral clustering. From an applicationsstandpoint, we employ spectral clustering to solve the so-called attritionproblem, where one aims to identify from a set of employees those who arelikely to voluntarily leave the company from those who are not. Our study shedslight on the empirical performance of existing approximate spectral clusteringmethods and shows the applicability of these methods in an important businessoptimization related problem.
展开▼
