#!/usr/bin/env python3
"""Generate a script that computes feature importance several ways on held-out data.
The script it writes loads a fitted model or Pipeline (joblib) and a HELD-OUT CSV (rows the model did not
train on), then computes, per input column:
gain impurity importance from the fitted model (LightGBM gain, XGBoost total_gain, CatBoost
PredictionValuesChange, sklearn feature_importances_), or coef_x_std (|coefficient| x
feature std) for linear models, mapped back from encoded columns to input columns by summing
permutation sklearn.inspection.permutation_importance on the whole pipeline with --repeats shuffles:
mean drop in the score and its std (so "indistinguishable from zero" can be judged)
group perm. joint permutation of clusters of correlated numeric features (|Spearman| >= --corr): the
importance of information that single-feature permutation hides when features share it
mean_abs_shap with --shap, mean |SHAP| from TreeExplainer (tree models), summed per input column
Held-out columns are reordered to the model's training order (CatBoost matches by position). It writes
importances.csv (wide: feature, gain or coef_x_std, permutation, permutation_std[, mean_abs_shap]) for
compare_importances.py, plus groups.csv with the grouped permutation results.
Needs scikit-learn, pandas, numpy, joblib (and shap with --shap).
Usage:
python gen_importance_script.py --model model.joblib --data holdout.csv --target churned -o importance.py
python gen_importance_script.py --model pipe.joblib --data test.csv --target price --scoring neg_root_mean_squared_error
python gen_importance_script.py --model m.joblib --data test.csv --target y --shap --repeats 20 --corr 0.8
Exit codes: 0 written, 2 usage error.
"""
import argparse
import os
import sys
TEMPLATE = '''"""Feature importance for {model!r} on held-out {data!r}. Generated by gen_importance_script.py.
Run: python {script}
Then: python compare_importances.py importances.csv
"""
import joblib
import numpy as np
import pandas as pd
from sklearn.inspection import permutation_importance
from sklearn.metrics import get_scorer
MODEL, DATA, TARGET, DROP = {model!r}, {data!r}, {target!r}, {drop!r}
SCORING, REPEATS, CORR, SHAP, SEED = {scoring!r}, {repeats}, {corr}, {shap}, 0
model = joblib.load(MODEL)
df = pd.read_csv(DATA)
df = df[df[TARGET].notna()]
X = df.drop(columns=[c for c in [TARGET] + DROP if c in df.columns])
y = df[TARGET]
trained = getattr(model, "feature_names_in_", None)
if trained is None:
trained = getattr(model, "feature_names_", None) # CatBoost matches columns by position
if trained is not None:
missing = [c for c in trained if c not in X.columns]
if missing:
raise SystemExit(f"held-out data lacks training columns: {{missing}}")
X = X[list(trained)]
if hasattr(model, "classes_") and y.dtype != object and set(np.unique(y)) - set(model.classes_):
raise SystemExit("target values do not match model.classes_")
cols = list(X.columns)
scorer = get_scorer(SCORING)
baseline = scorer(model, X, y)
print(f"held-out {{SCORING}}: {{baseline:.5f}} on {{len(X)}} rows")
steps = getattr(model, "steps", None)
est = model[-1] if steps else model
names = list(model[:-1].get_feature_names_out()) if steps else cols
def to_input(name):
"""Map an encoded column name (e.g. cat__plan_basic) back to its input column."""
short = name.split("__", 1)[-1]
for c in sorted(cols, key=len, reverse=True):
if short == c or short.startswith(c + "_"):
return c
return short
def per_input(values):
out = dict.fromkeys(cols, 0.0)
for n, v in zip(names, values):
key = to_input(n)
out[key] = out.get(key, 0.0) + float(v)
return out
# 1) impurity / model-based importance
kind = type(est).__module__.split(".")[0]
label = "gain"
if kind == "lightgbm":
raw = est.booster_.feature_importance(importance_type="gain")
elif kind == "xgboost":
score = est.get_booster().get_score(importance_type="total_gain")
fn = est.get_booster().feature_names or [f"f{{i}}" for i in range(len(names))]
raw = [score.get(f, 0.0) for f in fn]
elif kind == "catboost":
raw = est.get_feature_importance()
elif hasattr(est, "feature_importances_"):
raw = est.feature_importances_
elif hasattr(est, "coef_"):
Xt = model[:-1].transform(X) if steps else X.to_numpy()
Xt = Xt.toarray() if hasattr(Xt, "toarray") else np.asarray(Xt, dtype=float)
raw = np.abs(np.atleast_2d(est.coef_)).mean(axis=0) * Xt.std(axis=0)
label = "coef_x_std" # |coefficient| x feature std: not an impurity measure
else:
raw = None
gain = per_input(raw) if raw is not None and len(raw) == len(names) else None
# 2) permutation importance on the whole pipeline (input columns)
perm = permutation_importance(model, X, y, scoring=SCORING, n_repeats=REPEATS, random_state=SEED, n_jobs=-1)
# 3) grouped permutation for clusters of correlated numeric features
num = [c for c in cols if pd.api.types.is_numeric_dtype(X[c])]
corr = X[num].corr(method="spearman").abs().fillna(0) if num else pd.DataFrame()
parent = {{c: c for c in num}}
def find(c):
while parent[c] != c:
parent[c] = parent[parent[c]]
c = parent[c]
return c
for i, a in enumerate(num):
for b in num[i + 1:]:
if corr.loc[a, b] >= CORR:
parent[find(a)] = find(b)
clusters = {{}}
for c in num:
clusters.setdefault(find(c), []).append(c)
rng = np.random.default_rng(SEED)
group_rows = []
for members in clusters.values():
if len(members) < 2:
continue
drops = []
for _ in range(REPEATS):
Xp = X.copy()
idx = rng.permutation(len(Xp))
Xp[members] = Xp[members].to_numpy()[idx] # permute the group jointly (keeps its internal structure)
drops.append(baseline - scorer(model, Xp, y))
single = sum(perm.importances_mean[cols.index(m)] for m in members)
group_rows.append({{"group": "+".join(members), "size": len(members), "group_permutation": np.mean(drops),
"group_permutation_std": np.std(drops), "sum_single_permutation": single}})
# 4) optional mean |SHAP|
shap_imp = None
if SHAP:
import shap
Xt = model[:-1].transform(X) if steps else X
Xt = Xt.toarray() if hasattr(Xt, "toarray") else Xt
sv = shap.TreeExplainer(est).shap_values(Xt)
sv = np.asarray(sv)
if sv.ndim == 3:
sv = sv[:, :, -1]
shap_imp = per_input(np.abs(sv).mean(axis=0))
out = pd.DataFrame({{"feature": cols}})
if gain is not None:
out[label] = [gain[c] for c in cols]
out["permutation"] = perm.importances_mean
out["permutation_std"] = perm.importances_std
if shap_imp is not None:
out["mean_abs_shap"] = [shap_imp[c] for c in cols]
out = out.sort_values("permutation", ascending=False)
out.to_csv("importances.csv", index=False)
pd.DataFrame(group_rows).to_csv("groups.csv", index=False)
print(out.round(5).to_string(index=False))
for g in group_rows:
print(f"group {{g['group']}}: joint permutation {{g['group_permutation']:.5f}} "
f"vs sum of single {{g['sum_single_permutation']:.5f}}")
print("wrote importances.csv and groups.csv")
'''
def main():
ap = argparse.ArgumentParser(description="Generate a feature importance script.", epilog=__doc__,
formatter_class=argparse.RawDescriptionHelpFormatter)
ap.add_argument("--model", required=True, help="joblib file with the fitted model or pipeline")
ap.add_argument("--data", required=True, help="held-out CSV (not used for training)")
ap.add_argument("--target", required=True)
ap.add_argument("--drop", nargs="*", default=[])
ap.add_argument("--scoring", default="roc_auc", help="sklearn scorer name (default roc_auc)")
ap.add_argument("--repeats", type=int, default=10)
ap.add_argument("--corr", type=float, default=0.7, help="|Spearman| threshold for correlated groups")
ap.add_argument("--shap", action="store_true", help="also compute mean |SHAP| (tree models)")
ap.add_argument("-o", "--output")
a = ap.parse_args()
if a.repeats < 2 or not 0 < a.corr <= 1:
print("error: --repeats must be >= 2 and 0 < --corr <= 1", file=sys.stderr)
return 2
text = TEMPLATE.format(model=a.model, data=a.data, target=a.target, drop=a.drop, scoring=a.scoring,
repeats=a.repeats, corr=a.corr, shap=a.shap,
script=os.path.basename(a.output or "importance.py"))
if a.output:
with open(a.output, "w", encoding="utf-8") as f:
f.write(text)
print(f"wrote {a.output}: {a.scoring}, {a.repeats} repeats, groups at |rho| >= {a.corr}"
+ (", with SHAP" if a.shap else ""), file=sys.stderr)
else:
sys.stdout.write(text)
return 0
if __name__ == "__main__":
sys.exit(main())
anomaly-detection-model-builder
Build and evaluate anomaly detection for tabular, sensor and transaction data with few or no labels: define the anomaly type (point, contextual, collective), choose detectors (robust z-score/MAD and IQR rules, Mahalanobis with robust covariance, k-NN distance, LOF, Isolation Forest, one-class SVM, PCA reconstruction / autoencoders, forecasting residuals for time series), combine them by rank, explain each alert by its deviating features, set thresholds from an alert budget or the score distribution, and evaluate with precision@k, PR-AUC and recall on known incidents, or detector agreement when no labels exist. Includes pure-Python detectors (robust z, Mahalanobis, kNN, Isolation Forest, ensemble) and an evaluation script.
catboost-model-builder
Build CatBoost models that use its strengths correctly: raw categorical columns as cat_features (ordered target statistics, no manual encoding), text_features, Pool construction with NaN-safe categories, ordered boosting and has_time for temporal data, symmetric (oblivious) trees, overfitting detector with eval_set and use_best_model, losses for classification, regression (Quantile, Poisson, Tweedie) and ranking (YetiRank), class weights, monotone constraints, feature importance (PredictionValuesChange, LossFunctionChange) and SHAP values, GPU training and model export. Plans cat/text/numeric columns from a CSV and generates a runnable training script, and checks parameter sets for the conflicts CatBoost rejects (iterations vs n_estimators, Bayesian bootstrap with subsample, multiple class-weight options, depth > 16, use_best_model without eval_set).
categorical-encoding-selector
Choose and implement the right encoding for every categorical column given the model family and the data: one-hot (with infrequent-level grouping and handle_unknown), ordinal with an explicit order, cross-fitted target encoding (scikit-learn TargetEncoder, CatBoost ordered statistics, category_encoders inside CV), frequency/count encoding, feature hashing, native categorical support in LightGBM, HistGradientBoosting, XGBoost and CatBoost, prefix levels for codes (zip3, NAICS-2), and embeddings. Measures cardinality, rare-level share, unseen-level rate on a time-ordered holdout and target signal per column, emits the ColumnTransformer, and quantifies target-encoding leakage (in-sample vs out-of-fold vs holdout AUC).
classification-model-selector
Choose the right classification algorithm for tabular, text or mixed data and prove the choice with a fair bake-off: logistic regression (L2, L1, elastic net), linear SVM, naive Bayes, k-NN, RBF SVM, decision trees, random forest, extra trees, HistGradientBoosting, LightGBM, XGBoost, CatBoost, explainable boosting machines (EBM/GAM), points scorecards and MLPs, matched to data size, width, categorical and high-cardinality share, missingness, class imbalance, text, and constraints such as interpretability, monotonic constraints, calibrated probabilities, latency budgets and tiny deployable models. Ranks candidates with starting hyperparameters and cautions, then generates a bake-off script running every candidate on identical CV folds with per-fold metrics, fit time and prediction latency.
clustering-algorithm-selector
Choose the right clustering algorithm and check that clustering makes sense at all: cluster tendency (Hopkins statistic with a correlated-features caveat, nearest-neighbour isolation), then k-means / MiniBatchKMeans, Gaussian mixtures with BIC, agglomerative (Ward, average, complete, single) with dendrograms, DBSCAN, HDBSCAN, OPTICS, spectral clustering, BIRCH, mean shift, k-prototypes and k-modes for mixed or categorical data, and Gower distance, matched to data size, expected shapes, noise, varying density, known k, soft memberships, hierarchy and streaming. Covers preprocessing (scaling, encoding, dimensionality, distance metrics), validation (internal indices, bootstrap stability, external labels, business profiling) and assigning new points.
collaborative-filtering-implementer
Implement collaborative filtering recommenders that beat popularity: item-based kNN with cosine, Jaccard or asymmetric similarity, shrinkage, neighbourhood size, positive-only or mean-centred rating signals and 'because you liked X' explanations; implicit ALS with confidence weighting, BPR pairwise ranking and the closed-form EASE model in plain NumPy, tuned on a temporal validation split by recall@k and NDCG@k and refit on all data; plus user-kNN, explicit matrix factorisation, the implicit/LightFM libraries, hyperparameter ranges, cold-start fallbacks, scaling with sparse matrices and approximate nearest neighbours. Produces top-k lists in the format of the recommendation evaluator.