Mining multiple time sequences through co-clustering algorithms for distributional data

This paper deals with the co-clustering of distributional data applied to multiple time sequences. The aims are: to get a double-partition of data into clusters of units and variables; to summarize the main concepts in the data through histogram prototypes; to overview the evolution over time of the monitored phenomenon. We extend the double k-means algorithm to handle distributional data by using the L2 Wasserstein distance for comparing distributions. Moreover, we adapt double k-means algorithm to compute optimal relevance weights associated with the variables