Document Type


Publication Date



Computer Sciences


Producing meaningful clusterings for graph data requires the user to provide some insight to the program which he or she may not have regarding the data. The standard clustering algorithm, K-means, requires the user to specify k, the number of clusters to be produced by the algorithm. This paper discusses the recursivePartition algorithm, a recursive alternative to K-means clustering. The input to recursivePartition asks the user to specify n, the maximum size of any cluster. Using maximum cluster size and spectral methods based on the Laplacian matrix of the graph, recursivePartition has demonstrated an ability to produce highly accurate clusters over a range of inputs, even producing an exact match of the true clusterings present in the data in multiple tests. recursivePartition is capable of producing highly accurate clusters with a robustness to user input which the K-means clustering algorithm cannot match.


Approved by: Michael Heroux, Robert Hesse, Michael Gass, James Schnepf, Anthony Cunningham