inspection

On the test set, with the metric you actually care about. Running it on training data measures memorisation instead.

from sklearn.inspection import permutation_importance
import numpy as np

r = permutation_importance(
    model, X_test, y_test,           # TEST, not train
    n_repeats=20,
    scoring="average_precision",     # the metric that matters here
    n_jobs=-1, random_state=42,
)

order = r.importances_mean.argsort()[::-1]
for i in order[:12]:
    mean, std = r.importances_mean[i], r.importances_std[i]
    flag = "" if mean - 2 * std > 0 else "   (not distinguishable from 0)"
    print(f"  {feature_names[i]:28} {mean:+.4f} +/- {std:.4f}{flag}")

# Features whose importance is within two standard deviations of zero
# are not doing measurable work. That is a defensible reason to drop
# them, and it reads well in a model document.
#
# Caveat: with two highly correlated features, permuting either alone
# does little because the other carries the signal, so BOTH look
# unimportant. Check correlations before concluding.

Shows whether the model learned something monotonic and sensible, or something jagged that will not survive contact with new data.

from sklearn.inspection import PartialDependenceDisplay
import matplotlib.pyplot as plt

fig, ax = plt.subplots(figsize=(11, 4))
PartialDependenceDisplay.from_estimator(
    model, X_test,
    features=["months_delinquent", "income", ("income", "age")],
    kind="both",             # 'average' plus individual lines (ICE)
    subsample=200,
    ax=ax,
)
plt.tight_layout()

# What to look for: does risk rise monotonically with months
# delinquent? If the curve wiggles, the model has fitted noise, and
# a monotonic constraint in the booster will both improve
# generalisation and make the model defensible.