Relational Analysis for Clustering Consensus
Mustapha Lebbah, Younès Bennani, Nistor Grozavu, Hamid Benh · Machine Learning · 2010
In this chapter, we formally defined the problem of clustering and we presented an original and new approach of fusion/ensemble/consensus/aggregation clustering. The main idea was to find a clustering (or partition) of observations that represents the best consensus between several other clustering related to the same data set. The goal of the proposed algorithm is the improvement of confidence in cluster assignments by evaluating a history of cluster assignments for each observation. If we compare our algorithm (or method) to some recent clustering algorithm, we can assert that, unlike these new algorithms, our method is scalable, linear, in memory use and computational time and can handle data represented as observations cross attributes or as similarity matrix. Our clustering method handles missing values without replacing them by values that could be very far away from the true ones. It also contains a preprocessing module that, among other processings, can compute how discriminant are the attributes measured on the observations to be clustered. Finally we verified the intuitive appeal of the proposed approach and we studied the behavior of our algorithm on real and synthetic heterogeneous data sets. We observed that the proposed method increases performance as more as iterations of the process are performed. Another advantage of our method is that, neither do we need to re-process the data; nor do we need to fix the same cluster numbers for each application or clustering algorithm. In the future, we would like to perform a more detailed analysis involving huger data set and investigating the collaborative clustering.