The Relational Push-Pull Model: A Generative Model for Relational Data Clustering

Abstract

Relational data clustering is a form of relational learn-ing that clusters data using the relational structure of data sets to guide the clustering. Many approaches for rela-tional clustering have been proposed recently. The com-mon assumption in much of this research is that relations have a binding tendency (Neville, Adler, & Jensen 2003; Bhattacharya & Getoor 2007; Taskar, Segal, & Koller 2001; Neville & Jensen 2006). That is, edges are assumed to ap-pear more frequently within clusters than between clusters. This binding quality may be too strong an assumption. Bhattacharya & Getoor (2007) acknowledge that it is possi-ble for a relation to provide “negative evidence, ” where the presence of a relation between two objects implies that the objects belong in different clusters. If most of the edges in a relational set provide negative evidence, then this set should