Search

Ensemble learning can improve predictive performance compared to the performance of any of its constituents alone, while keeping computational demands manageable. However, no reference methodology is available for developing ensemble systems. In this paper, we adapt an ensemble framework for patent classification to assist data scientists in creating flexible ensemble architectures for text classification by selecting a finite set of constituent base models from the many available alternatives. We analyze the axes along which someone can select base models of an ensemble system and propose a methodology for combining them. Moreover, we conduct experiments to compare the effectiveness of ensemble systems against base models and state-of-the-art methods on multiple datasets (three patent classification and two text classification datasets), including long and short texts and single- and/or multi-labeled texts. The results verify the generality of our framework and the effectiveness of ensemble systems, especially ensembles of classifiers trained on different data sections/metadata. [ABSTRACT FROM AUTHOR]