EvalML
Most AutoML libraries optimise accuracy or log loss, and then leave you to discover that the highest-scoring model is the one that costs your business the most money. EvalML starts from the other end: you say what you are actually trying to win, and the search looks for the pipeline that wins it.
Objectives as first-class citizens
EvalML is an open-source AutoML library from Alteryx that builds and tunes complete pipelines — imputation, encoding, feature selection, estimator — for binary, multiclass, regression and time-series problems. Its distinguishing feature is the objective function. Alongside standard metrics it ships cost-based objectives such as LeadScoring and FraudCost, which take the value of a true positive and the price of a false one, and you can subclass the objective base class to encode your own economics. The search optimises that objective, and the decision threshold on binary problems is tuned against it too.
from evalml.automl import AutoMLSearch
automl = AutoMLSearch(X_train, y_train, problem_type='binary', objective='f1')
automl.search()
best = automl.best_pipelineWhat comes with it
Two surrounding pieces matter in practice. Data checks run before the search and flag the faults that quietly ruin unattended AutoML — target leakage, near-empty columns, mislabelled types, extreme class imbalance. Model-understanding tools (permutation importance, partial dependence, precision-recall and confusion-matrix plots) then make the winning pipeline explainable rather than merely delivered. EvalML sits in the same stack as Featuretools and Woodwork, so generated features, typed data and automated search compose into one workflow.
EvalML is an open-source AutoML library from Alteryx that builds, tunes and evaluates complete machine-learning pipelines for classification, regression and time-series problems. Its distinguishing feature is optimisation against custom objective functions, including cost-based ones such as fraud loss or lead value, rather than a generic accuracy metric. Built-in data checks catch target leakage and bad types before the search starts, which is where unattended AutoML most often goes wrong.
Ask a machine to find the best model and you first have to tell it what best means. Get that wrong and you get something that looks brilliant on paper and loses money in practice — a fraud detector, say, that catches nearly every cheat but blocks so many honest customers that it isn't worth running.
EvalML is a free tool that builds models for you and lets you state the goal in your own terms: what catching a real fraud is worth, what a false alarm costs. It then tries many combinations and keeps whichever scores best by that measure. Before it starts, it also inspects your data for the usual traps.