Model-Based Co-clustering for Functional Data
Abstract
In order to provide a simplified representation of key performance indicators for an easier analysis by mobile network maintainers, a model-based co-clustering algorithm for functional data is proposed. Co-clustering aims to identify block patterns in a data set from a simultaneous clustering of rows and columns. The algorithm relies on the latent block model in which each curve is identified by its functional principal components that are modeled by a multivariate Gaussian distribution whose parameters are block-specific. These latter are estimated by a stochastic EM algorithm embedding a Gibbs sampling. In order to select the numbers of row-and column-clusters, an ICL-BIC criterion is introduced. In addition to be the first co-clustering algorithm for functional data, the advantage of the proposed model is its ability to extract the hidden double structure induced by the data and its ability to deal with missing values. The model has proven its efficiency on simulated data and on a real data application that helps to optimize the topology of 4G mobile networks.
Origin | Files produced by the author(s) |
---|