Ensemble learning, social choice and collective intelligence: An experimental comparison of aggregation techniques

Ensemble learning provides a theoretically well-founded approach to address the bias-variance trade-off by combining many learners to obtain an aggregated model with reduced bias or variance. This same idea of extracting knowledge from the predictions or choices of individuals has been also studied under different perspectives in the domains of social choice theory and collective intelligence. Despite this similarity, there has been little research comparing and relating the aggregation strategies proposed in these different domains. In this article, we aim to bridge the gap between these disciplines by means of an experimental evaluation, done on a set of standard datasets, of different aggregation criteria in the context of the training of ensembles of decision trees. We show that a social-science method known as surprisingly popular decision and the three-way reduction, achieve the best performance, while both bagging and boosting outperform social choice-based Borda and Copeland methods.