AutoML

An AutoML system searches combinations of data transformations, model families and hyperparameters, sometimes building an ensemble from candidate models. Its scope depends on the system: some handle structured prediction, others neural architecture or specialized tasks. Search is driven by an evaluation metric and validation procedure, which the user must align with intended use. Automation explores a predefined space rather than discovering the correct target or causal question. It also inherits the assumptions of its components. A high-scoring selected pipeline can be complex, resource-intensive or inappropriate for the environment in which predictions must be served.

Define the target, allowed features and prediction-time boundaries before running a search. Configure group-aware or temporal validation when required and include resource, latency or interpretability constraints. Compare with a simple manually built baseline, inspect the chosen preprocessing and assess errors on a separate final test. Record package versions and export the full pipeline rather than only its estimator. The result should explain what the automation searched and why the chosen pipeline is acceptable, including any operational limits that were evaluated outside the search objective.

AutoML cannot repair a poorly defined target or unrepresentative labels. A default random split may be invalid for repeated entities or time-dependent data. Search can overfit validation, and a large ensemble may be difficult to explain or serve. Compatibility and export behavior depend on the implementation. Automation is broader than hyperparameter tuning when it also selects preprocessing or model families, but neither process yields unbiased final performance without an evaluation boundary that the search has not repeatedly inspected.