Model-Based Clustering for Diverse Data

Abstract: Clustering aims to group data into homogeneous subsets. For example, asteroids can be grouped into families originating from the same parent body, or stars placed in different spectral classes. This grouping is based on measurements collected for each object; these measurements may be of different types and are subsequently analyzed using a clustering method.

In this presentation, we focus on model-based clustering, which consists of modeling the overall data distribution as a mixture of several groups, each characterized by its own specific probability distribution. This approach embeds the clustering problem within the classical framework of inferential statistics, allowing the use of parameter estimation algorithms as well as model selection criteria.

After recalling the principles of model-based clustering, we will present several recent methodological developments. These include applications to the classification of temporal data, the identification of multiple clustering structures within a single dataset, and the classification of data based on recurrent events.