Tuning a Churn-Prediction Random Forest with GASearchCV

Description and use

GASearchCV boosted churn‑prediction recall from 48% to 83%, improving F1 from 0.543 to 0.604.

Problem

Improve recall of a churn‑prediction model for a phone company where missing a churner is costlier than a false alarm.

Outcome

Recall increased from 48% to 83% (≈+35 percentage points), precision dropped to 48%, but overall F1 improved from 0.543 to 0.604, demonstrating a net performance gain by optimizing for recall.

Workflow

• Define hyperparameter search space with Integer, Continuous, and Categorical ranges.
• Instantiate GASearchCV with 5‑fold CV, recall scoring, population_size=20, generations=15, n_jobs=-1.
• Fit GASearchCV on training data.
• Retrieve best hyperparameters and evaluate on test set.
• Compare recall, precision, and F1 to baseline.

Evidence

Recall (Churn) 0.481 (48%)

Source

Recall (Churn) 0.826 (83%)

Source

scoring='recall'

Source