Meta-Features (AutoML)
Before an AutoML system trains a single model, it can already say useful things about your data: how many rows it has, how lopsided the classes are, how much of it is text. Those measurements are meta-features — a description of the dataset rather than anything inside it.
What gets measured
A meta-feature vector mixes several kinds of fact. Simple counts cover rows, columns, and the ratio between them. Statistical descriptors cover skewness, kurtosis, and correlation between features. Information-theoretic ones cover class entropy and mutual information. Landmarkers are a shortcut of a different kind: train something deliberately cheap — a decision stump, a one-nearest-neighbour classifier — and record its accuracy, because how a trivial model fares on a dataset says a great deal about what a serious one will need.
What they are for
- Warm starts — Auto-sklearn computes meta-features for your dataset, finds the nearest datasets in its stored collection, and seeds Bayesian optimisation with the configurations that won there.
- Algorithm selection — a model trained on past runs maps meta-features to the algorithm likely to win, producing a shortlist before any training.
- Budget decisions — size and dimensionality tell the system whether full cross-validation is affordable or whether to subsample.
Why it matters
Without meta-features every run starts from nothing, and the first hour goes on rediscovering that gradient boosting does well on wide tabular data. They are what lets a search exploit the thousands of experiments that came before it — the dataset and run metadata published on OpenML exists largely to make that possible.
Meta-features are measurable properties of a dataset itself — row and column counts, class imbalance, feature skew, class entropy, and landmarker scores from deliberately cheap models. They describe the problem rather than solve it. They matter because they let an AutoML system recognise that a new dataset resembles ones it has already seen, and warm-start its search with configurations that worked there instead of exploring from zero.
A doctor glances at your chart — age, weight, blood pressure — before ordering any test. None of those numbers is a diagnosis, but together they narrow down what is worth checking first.
Meta-features are that chart, for a dataset. They are simple facts about the data: how big it is, how many columns it has, whether one category hugely outnumbers the others, how messy the numbers look. Automated machine-learning tools read these facts first, compare them against datasets they have handled before, and begin with the approaches that worked on similar data. It saves them from trying everything blindly.