Automated Ensembling
When a search has trained two hundred models, keeping only the winner is a strange way to spend the compute. The runners-up were wrong in different places, and blending their opinions beats the single best model more reliably than further tuning would. Automated ensembling is the step that gathers those votes without being asked.
How it works
Every candidate trained during the run is kept along with its predictions on held-out folds. A combiner then searches for weights over that library, scored on the same validation data used for selection:
- Greedy ensemble selection repeatedly adds the model that most improves the validation score, with replacement — so a strong member can be added twice and earn a heavier weight.
- Stacking trains a second-level model on out-of-fold predictions, learning which base model to trust in which region of the data.
- Bagged refits retrain the chosen members on all the data once the weights are fixed.
Why it matters
Most of the margin on an AutoML leaderboard comes from this stage, which is why auto-sklearn, H2O AutoML and AutoGluon all end a run with an ensemble rather than a champion. The cost is honesty about complexity: you ship twenty models instead of one, inference gets slower, and explanation gets harder. And when weights are fitted on the same folds that selected the members, the reported score drifts optimistic — which is why a final untouched test set still earns its keep.
Automated ensembling combines the many models produced during an AutoML search into one predictor, with membership and weights chosen automatically. Greedy ensemble selection or stacking over out-of-fold predictions turns a library of candidates — including mediocre ones that err differently — into a model stronger than the single best. It matters because it recovers value from every configuration the search trained, and it supplies most of the accuracy gap between an AutoML run and one hand-tuned model.
Ask one doctor for a diagnosis and you get one opinion. Ask five and look for the pattern in their answers, and you usually do better, because their individual blind spots cancel out.
A machine-learning system that searches for a good model trains dozens of them along the way. Automated ensembling is the habit of keeping the whole crowd instead of only the champion. The software works out how much to trust each model and blends their answers into one prediction. A mistake that only one model makes gets outvoted, which is why a committee put together this way beats the best individual member — and why the time spent on the also-rans is not wasted.