Unsupervised ranking and characterization of differentiated clusters

We describe a framework for automatically identifying and visualizing the most differentiating attributes of each cluster in a clustered data set. A dissimilarity function measures the cluster-conditional distinguishing saliency of each attribute with respect to a reference realization of the same attribute. For each cluster, the N attributes that are most dissimilar are presented first to the human expert, along with the overall dissimilarity of the cluster. We discuss the computational benefits of the proposed framework, how it can be implemented with map-reduce, its application to the behavioral analysis of mobile phone users, and it broad applicability to diverse problem domains.