Section: New Results

Axis 1: Co-clustering: A versatile way to perform clustering

Participant : Christophe Biernacki.

Standard model-based clustering is known to be very efficient for low dimensional data sets, but it fails for properly addressing high dimension (HD) ones, where it suffers from both statistical and computational drawbacks. In order to counterbalance this curse of dimensionality, some proposals have been made to take into account redundancy and features utility, but related models are not suitable for too many variables. We advocate that the latent bloc model, a probabilistic model for co-clustering, is of particular interest to perform HD clustering of individuals even if it is not its primary function. We illustrate in an empirical manner the trade-off bias-variance of the co-clustering strategy in scenarii involving HD fundaments (correlated variables, irrelevant variables) and show the ability of co-clustering to outperform simple mixture row-clustering. An early version of this work has been presented to an national conference with international audience [46].

We also co-organized a special session to an international conference [45] to discuss the potential links between deterministic methods for co-clustering (based on a metric and computer science procedure) or probabilistic methods for co-clustering (mainly based on mixture models). It was the opportunity to gather related communities which are often distinct.

All are joint works with Christine Keribin from Université Paris-Sud.