Boosting Unsupervised Additive Clustering Using Cluster-Wise Optimization and Multi-label Learning

Additive or overlapping clustering is a technique that is used to analyze overlapping cluster structure in data. In this paper, we motivate the overlapping clustering problem using an example of categorizing movies. We describe the ADCLUS and INDCLUS overlapping clustering models as discrete versions of the CANDECOMP/PARAFAC models. We describe the scalability problems inherent in current overlapping clustering approaches. We give a framework and algorithm for scaling up unsupervised overlapping clustering using a combination of a cluster-wise optimization technique and techniques from multi-label learning. Our framework uses a subset of data to find a training solution and then uses multi-label techniques to find labels for the remaining data.