Attribute Clustering Based on Heuristic Tree Partition

Attribute clustering has been previously employed to detect statistical dependence between subsets of variables. Clusters of variables can be appropriately used for detecting highly dependent domain variables and then reducing the complexity of learning Bayesian networks. We propose a novel attribute clustering algorithm motivated by research of complex networks, called the Star Discovery algorithm. The algorithm partitions and indirectly discards inconsistent edges from a maximum spanning tree by starting appropriate initial modes, therefore generating stable clusters. It discovers sound clusters through simple graph operations and achieves significant computational savings. We compare the Star Discovery algorithm against earlier attribute clustering algorithms and evaluate the performance in several domains.