A General and Scalable Approach to Mixed Membership Clustering

Spectral clustering methods are elegant and effective graph-based node clustering methods, but they do not allow mixed membership clustering. We describe an approach that first transforms the data from a node-centric representation to an edge-centric one, and then use this representation to define a scalable and competitive mixed membership alternative to spectral clustering methods. Experimental results show the proposed approach improves substantially in mixed membership clustering tasks over node clustering methods.