How can I reduce model training time?

Use Fast run mode for early iteration, begin with the smallest useful graph, and expand the data and search space only after the workflow is correct.

Choose the run mode

• Fast reduces the AutoML search and is approximately four times faster than Normal.
• Normal balances search time and model quality.
• Best expands the search and is approximately four times more thorough than Normal.

Changing the run mode changes the search budget, so faster training can trade off against finding a stronger configuration. See Run Mode and Model Plan.

Confirm the selected mode in your workflow. The AutoML interface and individual SDK methods can define different defaults.

Start with a smaller graph

For the first training run, include the prediction entity, target data, and the minimum set of related tables needed to test the task. Confirm the following before adding more data:

• Labels and class distributions match expectations.
• Training, validation, and holdout splits cover the intended periods.
• The graph contains the expected relationships.
• No input column reveals the prediction target.

Add tables and columns incrementally, measuring whether each addition improves validation or holdout performance.

Limit the development history

Using a shorter, representative time range can reduce early iteration cost. Preserve enough history to include the behavior and seasonality relevant to the prediction task. Restore the intended production history before final evaluation.

Downsample relational data carefully

If source-side sampling is necessary, sample by the prediction entity and retain its related rows across every table. Do not independently select a percentage of rows from each table, because that can break relationships and remove entity histories.

See Data Preparation Best Practices and Training for the surrounding workflow.