ordinalClust: An R Package to Analyze Ordinal Data

International audienceOrdinal data are used in many domains, especially when measurements are collectedfrom people through observations, tests, or questionnaires. ordinalClust is an innovative R packagededicated to ordinal data that provides tools for modeling, clustering, co-clustering and classifyingsuch data. Ordinal data are modeled using the BOS distribution, which is a model with two meaningfulparameters referred to as "position" and "precision". The former indicates the mode of the distributionand the latter describes how scattered the data are around the mode: the user is able to easily interpretthe distribution of their data when given these two parameters. The package is based on the coclustering framework (when rows and columns are simultaneously clustered). The co-clusteringapproach uses the Latent Block Model (LBM) and the SEM-Gibbs algorithm for parameter inference.On the other hand, the clustering and the classification methods follow on from simplified versionsof the SEM-Gibbs algorithm. For the classification process, two approaches are proposed. In thefirst one, the BOS parameters are estimated from the training dataset in the conventional way. In thesecond approach, parsimony is introduced by estimating the parameters and column-clusters from thetraining dataset. We empirically show that this approach can yield better results. For the clusteringand co-clustering processes, the ICL-BIC criterion is used for model selection purposes. An overviewof these methods is given, and the way to use them with the ordinalClust package is described usingreal datasets. The latest stable package version is available on the Comprehensive R Archive Network(CRAN)