Automated Pipeline Configuration

A model is the last step of a pipeline, not the whole of it. Before it sees anything, missing values get filled, categories encoded, features scaled or dropped, classes rebalanced — and every one of those steps has settings that interact with the model's own. Automated pipeline configuration is the search over that entire assembly.

What the search ranges over

Which steps exist at all — whether to select features, whether to reduce dimensions, whether to resample the classes.
Which operator fills each slot — median imputation or iterative, one-hot or target encoding, PCA or nothing.
The hyperparameters of whatever was chosen — a conditional space, since PCA's component count only exists when PCA is in the pipeline.

Why the search has to be joint

Tune the preprocessing against one model, then swap the model, and the preprocessing is now tuned for the wrong thing. Scaling is irrelevant to a decision tree and decisive for nearest neighbours; aggressive feature selection that rescues a linear model starves a boosted ensemble. Only a search over steps and model together sees those interactions.

In practice

TPOT evolves pipeline graphs with genetic programming, auto-sklearn optimises a fixed-shape scikit-learn pipeline with the Bayesian optimiser SMAC, and FLAML trades accuracy against cost under a time budget. Each returns a plain Pipeline object you can read, version and refit — the configuration is an inspectable artefact, not a black box.

Automated pipeline configuration is the automatic choice of which preprocessing steps a model needs, which operator fills each step, and how all of them are parameterised. The space is conditional — an operator's hyperparameters exist only when that operator is selected — and it is searched jointly with the model, because preprocessing tuned for one learner suits another badly. Without it, the steps around the model sit at defaults nobody validated.

Before data reaches the part that does the learning, it travels down a short assembly line: fill in the blanks, turn words into numbers, put everything on a comparable scale, perhaps drop the columns that add nothing. Each station on that line has choices, and those choices affect one another.

Automated pipeline configuration means the software designs the assembly line as well as the model at the end of it. It tries different stations, in different arrangements, with different settings, and keeps whichever complete line predicts best. What you get back is a full recipe you can run again on fresh data — not just a model that happened to work once.