The aim of collaborative clustering is to enhance the performances of clustering algorithms by enabling them to work together and exchange their information to tackle difficult data sets. The fundamental concept of collaboration is that clustering algorithms operate locally but collaborate by exchanging information about the local structures found by each algorithm. This kind of collaborative learning can be beneficial to a wide number of tasks including multi-view clustering, clustering of distributed data with privacy constraints, multi-expert clustering and multi-scale analysis. Within this context, the main difficulty of collaborative clustering is to determine how to weight the influence of the different clustering methods with the goal of maximizing the final results and minimizing the risk of negative collaborations—where the results are worse after collaboration than before. In this paper, we study how the quality and diversity of the different collaborators, but also the stability of the partitions can influence the final results. We propose both a theoretical analysis based on mathematical optimization, and a second study based on empirical results. Our findings show that on the one hand, in the absence of a clear criterion to optimize, a low diversity pool of solution with a high stability are the best option to ensure good performances. And on the other hand, if there is a known criterion to maximize, it is best to rely on a higher diversity pool of solution with a high quality on the said criterion. While our approach focuses on entropy based collaborative clustering, we believe that most of our results could be extended to other collaborative algorithms.
This is an open access article distributed under the Creative Commons Attribution License
which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited