Ensembles for Predicting Structured Outputs

While ensembles have been used for structured output learning, the literature lacks an extensive study of different strategies to construct ensembles in this context. In this work, we fill this gap by presenting a thorough empirical comparison of ensembles that predict the complete output structure at once, versus a combination of ensembles that each predicts a single component of the structure. We present results in two structured output learning tasks, using predictive clustering trees as base learners. The main results are that the difference in predictive performance is not significantly different for both approaches. However, in terms of total model size and induction times, ensembles that exploit the output structure are significantly more efficient.