Clustering Procedures and Module Detection

Abstract

Detecting clusters (also referred to as groups or modules) of closely related objects is an important problem in data mining in general. Network modules are often defined as clusters. Partitioning-around-medoids (PAM) clustering and hierarchical clustering are often used in network applications. Partitioning-around-medoids (aka. k-medoid clustering) leads to relatively robust clusters but requires that the user specify the number k of clusters. Hierarchical clustering is attractive in network applications since (a) it does not require the specification of the number of clusters and (b) it works well when there are many singleton clusters and when cluster sizes vary greatly. But hierarchical clustering requires the user to determine how to cut branches of the resulting cluster tree. Toward this end, one can use the dynamicTreeCut method and R library. The dynamic hybrid method combines the advantages of hierarchical clustering and partitioning-around-medoids clustering. Network concepts are useful for defining cluster quality statistics (e.g., to measure the density or separability of clusters). To determine whether the cluster structure is preserved in another data sets, one can use cross-tabulation-based preservation statistics. To measure the agreement between two clusterings, one can use the Rand index and other cross-tabulation-based statistics.