Hyperparameter Optimization (Engineering View)

Role

Optuna is used to search LightGBM hyperparameters (and optionally feature set) for detection, multiclass, or regression pipelines. The objective is validation metric (or cross-validation over cases), so the chosen configuration tends to generalize.

---

Why It Matters for Engineering

• Generalization: Tuning on validation (with by-case split) reduces overfitting (see Overfitting analysis and Reducing overfitting).
• Consistency: The same data loading, schema, and split logic as in training ensure that the optimized params are directly usable in the training scripts.
• Optional feature selection: When “pre–feature selection” is enabled, Optuna runs after a feature-selection run so the search space is over a reduced schema.

---

Configuration and Pipelines

• Config: Same section as the corresponding training pipeline; plus n_trials, cv_folds, timeout, enable_feature_selection_before_tuning, and constraints (e.g. min_recall).
• Script: run_optimize_lgbm_hyperparameters.py.
• Detailed reference: See Hyperparameter optimization in the main Pipelines section.

This Engineering page gives the rationale; the full how-to and options are in the Pipelines docs.