Which Way, and How Much?

A clinician asked for the model's top columns, and for each one, which way it pushes the prediction. That needs two tools: importance for which, partial dependence for which way.

Task: write top_with_direction(n).

• Load load_diabetes(return_X_y=True, as_frame=True) and split with train_test_split(X, y, test_size=0.25, random_state=0).
• Fit make_pipeline(StandardScaler(), LinearRegression()) on the training rows.
• Rank the columns by permutation_importance(model, X_test, y_test, n_repeats=10, random_state=0), highest mean first.
• For each of the top n columns, run partial_dependence(model, X_train, features=[name], grid_resolution=10) and take its ["average"][0]. The column rises if the last value is greater than the first, and falls otherwise.
• Return a list of [name, "rises" or "falls"] pairs, in importance order.