Auto-Stacker
Stacking — training a second model to combine the predictions of several others — reliably squeezes out extra accuracy, and it is fiddly enough that most people skip it. An auto-stacker is the part of an AutoML system that does it for you.
What it automates
A hand-built stack needs a decision at every layer: which base models go in, how many layers deep to go, how to produce the out-of-fold predictions the next layer trains on — leak those and the stack validates beautifully then fails in production — and which model sits on top. An auto-stacker makes those choices from data: it draws base learners from the models already trained during search, generates the next layer's features by k-fold bagging, fits the meta-learner, and validates the whole structure with no human in the loop. The name also belongs to a specific research system, Autostacker, which evolved layered pipelines where each layer's predictions are appended to the original features as synthetic columns for the layer above.
Where you meet it
- AutoGluon builds multi-layer stack ensembles with repeated bagging by default — the main reason its best-quality preset beats a single tuned model.
- H2O AutoML closes every run by fitting two stacked ensembles: one over all models, one over the best of each family.
- Auto-sklearn takes the lighter route of greedy ensemble selection — weighted voting rather than a trained meta-learner.
Automated stacking is where much of AutoML's measured advantage over careful manual modelling comes from: the ensemble gets assembled correctly every time, including the out-of-fold discipline a human working to a deadline tends to cut.
An auto-stacker is the component of an AutoML system that builds a stacked ensemble automatically: it selects base models, generates leak-free out-of-fold predictions from them, and trains a meta-model on those predictions. Assembling a stack by hand is error-prone, and a stack usually outperforms any single tuned model, so automated stacking is how frameworks such as AutoGluon and H2O AutoML reach their strongest results.
When a hard decision has to be made, one sensible move is to ask several experts and then have someone experienced weigh up their answers — knowing that one is usually too optimistic and another is only good with certain kinds of cases.
That final weighing-up is what stacking does with machine-learning models, and an auto-stacker is software that assembles the whole panel by itself. It decides which models to ask, works out how much to trust each one, and blends them into a single answer. The result is normally more accurate than any one model on its own, and nobody had to set it up by hand or know which combination would work.