A Parameter-Free Method for Discovering Generalized Clusters in a Network

We show that an MDL-based graph clustering method may be used for discovering generalized clusters from a graph and then extend it so that the input is a network. We define intuitively that generalized clusters contain at least a cluster in which nodes are connected sparsely and the cluster is connected either densely to another cluster or sparsely to another conventional cluster. The first characteristic of the MDL-based graph clustering is a direct outcome of an entropy function used in measuring the encoding length of clusters and the second one is realized through our new encoding method. Experiments using synthetic and real data sets give promising results.