Demo: model explainability

Two modern tree models (XGBoost and CatBoost) were trained to predict customer churn. Compare explanation methods, measure the quality of each one and see which one the demo suggests.

1. Which variables matter most?

Each method answers differently. The first column is the reference: the real effect measured by “removing” each variable from the model. Variables marked as noise were created with no effect at all.

2. Which method explains best?

Each method gets scores from 0 to 1 on five criteria. Adjust how much each criterion matters and watch the suggestion change.

Reset weights

3. Explaining one customer

SHAP and LIME explain the same case. Bars to the right increase churn risk; to the left, they reduce it.