Rethinking Collaborative Clustering: A Practical and Theoretical Study Within the Realm of Multi-view Clustering

With distributed and multi-view data being more and more ubiquitous, the last 20 years have seen a surge in the development of new multi-view methods. In unsupervised learning, these are usually classified under the paradigm of multi-view clustering: A broad family of clustering algorithms that tackle data from multiple sources with various goals and constraints. Methods known as collaborative clustering algorithms are also a part of this family. Whereas other multi-view algorithms produce a unique consensus solution based on the properties of the local views, collaborative clustering algorithms aim to adapt the local algorithms so that they can exchange information and improve their local solutions during the multi-view phase, but still produce their own distinct local solutions. In this chapter, we study the connections that collaborative clustering shares with both multi-view clustering and unsupervised ensemble learning. We do so by addressing both practical and theoretical aspects: First we address the formal definition of what is collaborative clustering as well as its practical applications. By doing so, we demonstrate that pretty much everything called collaborative clustering in the literature is either a specific case of multi-view clustering, or misnamed unsupervised ensemble learning. Then, we address the properties of collaborative clustering methods, and in particular we adapt the notion of clustering stability and propose a bound for collaborative clustering methods. Finally, we discuss how some of the properties of collaborative clustering studied in this chapter can be adapted to broader contexts of multi-view clustering and unsupervised ensemble learning.