Variants of the Fast Adaptive Stacking of Ensembles algorithm

Abstract

The treatment of large data streams in the presence of concept drifts is one of the main challenges in the fields of machine learning and data mining. This dissertation presents two families of ensemble algorithms designed to quickly adapt to concept drifts, both abrupt and gradual. The families Fast Stacking of Ensembles boosting the Old (FASEO) and Fast Stacking of Ensembles boosting the Best (FASEB) are adaptations of the Fast Adaptive Stacking of Ensembles (FASE) algorithm, designed to improve run-time and memory requirements, without presenting a significant decrease in terms of accuracy when compared to the original FASE. In order to achieve a more efficient model, adjustments were made in the update strategy and voting procedure of the ensemble. To evaluate the proposals against state of the art methods, Naïve Bayes (NB) and Hoeffding Tree (HT) are used, as learners, to compare the performance of the algorithms on artificial and realworld data-sets. An extensive experimental investigation with a total of 70 experiments and application of Friedman and Nemenyi statistical tests showed the families FASEO and FASEB are more efficient than FASE with respect to execution time and memory in many scenarios, often also achieving better accuracy results.