How do I tune and improve a model?

Diagnose the Results

Check the target, entity population, and prediction horizon. Inspect labels, missing values, possible leakage, graph relationships, and entity history. Review validation results for the cohorts where performance is weakest.

Use Training Diagnostics and Model Explainability to choose what to investigate. Add data or change settings one at a time, measuring each change on validation data. Keep the final test set separate from model selection.

Adjust the Model Plan

Call query.suggest_model_plan(snapshot), inspect the returned plan, and change its settings before fitting a Trainer. The plan controls experiment count, encoders, neighbor sampling, architecture, optimization, and evaluation metrics.

Labels and split assignments belong to the separate TrainingTableGenerationPlan used to generate training rows. See the Model Plan Guide for supported model settings.

Control Training Time

For early iterations, request a plan with run_mode=RunMode.FAST, reduce plan.training_job.num_experiments and plan.optimization.max_epochs, and disable refits if you do not need them. See Training for imports and example settings.

Use the tables and history needed for the task. If you sample source data, retain related rows across tables so histories and relationships remain intact. Compare runtime and validation quality before reducing the budget further.