Meta-Model Selection
Faced with a new dataset you could train every candidate algorithm and keep the winner. Meta-model selection takes a shortcut: use a model to predict which model will win, trained on what happened the last few thousand times.
How the prediction is built
The training data is a table of past experiments — one row per dataset, described by its meta-features, holding the performance each algorithm achieved on it. A learner is fitted over that table, and depending on the framing it predicts an outright winner (classification), each candidate's score (regression), or an ordering (learning to rank). Because the prediction concerns algorithms rather than rows, the same problem is called algorithm selection, and its classic framing — map a problem instance to the best solver for it — predates machine learning entirely.
How it is used
- Seeding a search — Auto-sklearn uses it to choose the first configurations it evaluates, so Bayesian optimisation starts somewhere sensible.
- Narrowing a portfolio — drop the families the meta-model is confident will lose, and spend the saved budget tuning the rest.
- Per-dataset choice — no single algorithm wins everywhere, which is exactly why a selector can beat always picking the overall favourite.
Why it matters
It turns the cost of choosing from a full training sweep into a single cheap inference, and the quality of that choice rests entirely on the breadth of experiment history behind it. Fed a large repository of past runs it is a strong prior; fed a dozen datasets it inherits their bias and recommends gradient boosting for everything.
Meta-model selection predicts which learning algorithm will perform best on a new dataset by training a model over the record of past experiments — datasets described by meta-features, paired with the performance each algorithm achieved. It replaces a full training sweep with a single inference, and is used to seed or prune an AutoML search. Its usefulness scales with the breadth of experiment history behind it; a narrow history yields confidently biased recommendations.
A good mechanic hears your engine and names the likely fault before opening anything. It is not a guess — they have listened to a few thousand engines, and the sound narrows it down.
Meta-model selection gives software that instinct. Rather than training every possible model on your data and comparing the results, the system looks at basic facts about your dataset, recalls what worked on similar ones, and predicts which approach will do best. It is a shortcut built out of experience, so it is only as good as the range of past jobs it has to draw on.