Mixing hetero- and homogeneous models in weighted ensembles

The effectiveness of ensembling for improving classification performance is well documented. Broadly speaking, ensemble design can be expressed as a spectrum where at one end a set of heterogeneous classifiers model the same data, and at the other homogeneous models derived from the same classification algorithm are diversified through data manipulation. The cross-validation accuracy weighted probabilistic ensemble is a heterogeneous weighted ensemble scheme that needs reliable estimates of error from its base classifiers. It estimates error through a cross-validation process, and raises the estimates to a power to accentuate differences. We study the effects of maintaining all models trained during cross-validation on the final ensemble's predictive performance, and the base model's and resulting ensembles' variance and robustness across datasets and resamples. We find that augmenting the ensemble through the retention of all models trained provides a consistent and significant improvement, despite reductions in the reliability of the base models' performance estimates.