Uberband : meta-aprendizado e otimização baseada em bandidos multi-armados para seleção eficiente e efetiva de processos completos

Abstract

Since data management and storage technologies become widely available, it becomes a challenge to provide users with effective systems for analyzing and understanding these data. Knowledge Discovery on Databases (KDD) is the non trivial process of extracting interesting, valid, and useful patterns from data. This process ranges from data selection to interpretation of the identified patterns. Especially for non-expert users, the definition and management of KDD process are complex activities, since it requires knowledge on how to choose the appropriate operators from the available range, how to configure them and how to interpret their output. Automatic Workflow Selection (AWS) aims to assist users of KDD in the onerous task of choosing the workflow, which includes preprocessing methods, machine learning algorithms and their hyper-parameter configurations, more suitable for a given problem. Although several solutions already exist for this task, such solutions are limited from the point of view of experimental evaluation of candidate workflows: i) some solutions do not perform workflow experimentation and are based on performance predictions in similar problems, which can lead to non-precise recommendations and ii) other solutions evaluate the workflows configurations over the entire training set until the best option is found. These latter solutions usually get more accurate results, however, they become computationally time-consuming as the datasets increase and new algorithms are developed. In this sense, this research proposes and investigates a new algorithm for AWS, named Uberband, that combines metalearning and multi-armed bandit optimization to perform adaptive allocation of the training data set during the optimization process. Results of the comparative experimental analysis with state-of-the-art solutions in AWS indicated that Uberband provides a AWS with good performance and in a significantly speedup over the current solutions.