Abstract : The internet and recent architectures such as sensor networks are currently witnessing tremendous and continuously growing amounts of data, often distributed on large scales. Combined with user expectations with respect to tooling, this encourages developing adequate techniques for analyzing and indexing. Classification and clustering tasks are about characterizing classes within data collections. These are often used as building blocks for designing tools aimed at making data accessible to users. In this document, we describe our contributions to mixture models aggregation. These models are classically used for content categorization. Using variational Bayesian principles, we aimed at designing low computation and transmission costs algorithms. Doing so, we aimed at proposing a building block for distributed density model estimation. We also contributed to visual classification applied to data streams. To this purpose, we employed bio-mimetic principles, and results from graph theory. More specifically, visual and dynamic abstractions of an underlying clustering process were proposed. We strived to provide users with efficient interfaces, while allowing using their actions as a feedback.