Clustering Relational Data Based on Randomized Propositionalization

Clustering of relational data has so far received a lot less attention than classification of such data. In this paper we investigate a simple approach based on randomized propositionalization, which allows for applying standard clustering algorithms like KMeans to multi-relational data. We describe how random rules are generated and then turned into boolean-valued features. Clustering generally is not straightforward to evaluate, but preliminary experimental results on a number of standard ILP datasets show promising results. Clusters generated without class information usually agree well with the true class labels of cluster members, i.e. class distributions inside clusters generally differ significantly from the global class distributions. The two-tiered algorithm described shows good scalability due to the randomized nature of the first step and the availability of efficient propositional clustering algorithms for the second step.