STARS: A new ensemble partitioning approach

In this work, we propose a novel ensemble learning approach based on a fast partitioning structure called STARS: Several Thresholds on a Random Subspace. Instead of modeling directly the posterior distribution over the entire space, we estimate an ensemble of posterior distributions in different random directions. This permits breaking down the complexity of learning distributions in high-dimensional spaces. By aggregating the predictions of multiple independent STARS elements, a strong multi-class ensemble can be constructed. Our approach can be instantiated for different tasks such as classification, clustering or regression, and this in an offline or online fashion. We show in the current paper the performance of our approach on several multi-class classification experiments on benchmark datasets. Furthermore, we instantiate STARS for clustering in the context of dictionary learning applied to image categorization and modality recognition of medical images.