Automated Machine Learning
Model selection, tuning and evaluation without a human in the loop — and the questions that still need one.
AutoML automates the search over models and hyperparameters, and it does it well. Given a clean tabular dataset, a target column and a metric, a modern AutoML system will usually match or beat a competent analyst working by hand, in far less time.
What the metric hides
The search optimises what you asked for. If the metric is wrong for the decision, the winner is confidently wrong. If the validation split leaks information, the score is inflated and nothing in the pipeline will notice.
Where AutoML is genuinely strong
Wide tabular data with many weak predictors, where gradient-boosted ensembles dominate and hand-tuning is drudgery. Rapid baselines. Systematic comparison across families that a person would not have the patience to run.
Where to be careful
Small datasets, where the search overfits the validation set. Time series, where naive cross-validation is invalid. Any setting where the model must be defended to someone who did not build it — see explainability.