#!/usr/bin/env python3
"""Generate a script that computes SHAP values correctly for a saved model and exports them as CSV.
The script it writes loads a joblib model (a bare estimator or a scikit-learn Pipeline) and a data CSV, then:
- for a Pipeline, transforms the data with all steps but the last and explains the final estimator on the
transformed columns (names from get_feature_names_out); with --group, sums the SHAP columns produced
from one input column (one-hot levels, splines) back into that column, which keeps additivity exact
- chooses the explainer from the model type at run time:
LightGBM / XGBoost / sklearn trees and forests -> shap.TreeExplainer
(--perturbation interventional with a background sample: SHAP relative to the background mean
prediction; tree_path_dependent: no background, uses the tree's cover statistics; always used for
HistGradientBoosting, whose interventional output in shap 0.54 misses the baseline)
CatBoost -> CatBoost's own get_feature_importance(type="ShapValues") (native categorical features)
linear / logistic -> shap.LinearExplainer with the background
anything else -> shap.PermutationExplainer on predict_proba / predict (model-agnostic, slower)
- explains the raw model output (log-odds for boosted and linear classifiers, probability for sklearn
forests and trees, the prediction for regressors) for tree and linear models, and the probability for the model-agnostic path; selects --class for multiclass
- writes shap_values.csv (features + base_value + prediction) and shap_features.csv (the explained
feature values) for shap_summary.py and explain_row.py, and asserts additivity
The generated script needs shap, pandas, numpy, joblib and the model's library.
Usage:
python gen_shap_script.py --model model.joblib --data test.csv --target churned -o explain.py
python gen_shap_script.py --model pipe.joblib --data test.csv --drop customer_id --group --rows 2000
python gen_shap_script.py --model rf.joblib --data x.csv --perturbation tree_path_dependent --class 2
Exit codes: 0 written, 2 usage error.
"""
import argparse
import os
import sys
TEMPLATE = '''"""Compute SHAP values for {model!r} on {data!r}. Generated by gen_shap_script.py.
Run: python {script}
Then: python shap_summary.py shap_values.csv --features shap_features.csv
python explain_row.py shap_values.csv --features shap_features.csv --row 0{link_hint}
"""
import joblib
import numpy as np
import pandas as pd
import shap
MODEL, DATA, TARGET, DROP = {model!r}, {data!r}, {target!r}, {drop!r}
ROWS, BACKGROUND, PERTURBATION, CLASS, GROUP, SEED = {rows}, {background}, {perturbation!r}, {cls}, {group}, 0
model = joblib.load(MODEL)
df = pd.read_csv(DATA)
X_raw = df.drop(columns=[c for c in [TARGET] + DROP if c and c in df.columns])
if len(X_raw) > ROWS:
X_raw = X_raw.sample(ROWS, random_state=SEED)
X_raw = X_raw.reset_index(drop=True)
steps = getattr(model, "steps", None)
if steps: # explain the last step on the transformed data
prep, est = model[:-1], model[-1]
X = prep.transform(X_raw)
X = X.toarray() if hasattr(X, "toarray") else np.asarray(X)
names = list(prep.get_feature_names_out())
X = pd.DataFrame(X, columns=names)
else:
prep, est, X = None, model, X_raw.copy()
names = list(X.columns)
kind = type(est).__module__.split(".")[0] + "." + type(est).__name__
is_clf = hasattr(est, "predict_proba")
background = X.sample(min(BACKGROUND, len(X)), random_state=SEED)
def pick(values, base):
values, base = np.asarray(values), np.atleast_1d(np.asarray(base, dtype=float))
if values.ndim == 3: # (rows, features, classes)
k = CLASS if CLASS is not None else values.shape[2] - 1
return values[:, :, k], float(base[k])
return values, float(base[-1] if base.size > 1 and CLASS is None else base[CLASS or 0] if base.size > 1 else base[0])
if kind.startswith("catboost."):
from catboost import Pool
cat = [i for i, c in enumerate(X.columns) if not pd.api.types.is_numeric_dtype(X[c])]
raw = est.get_feature_importance(Pool(X, cat_features=cat or None), type="ShapValues")
if raw.ndim == 3: # (rows, classes, features + 1)
raw = raw[:, CLASS if CLASS is not None else -1, :]
sv, base = raw[:, :-1], float(raw[0, -1])
output = est.predict(X, prediction_type="RawFormulaVal")
output = output[:, CLASS if CLASS is not None else -1] if np.ndim(output) == 2 else output
space = "raw (log-odds for classifiers)" if is_clf else "prediction"
elif kind.split(".")[0] in ("lightgbm", "xgboost") or "Tree" in kind or "Forest" in kind or "Boosting" in kind:
pert = PERTURBATION
if "HistGradientBoosting" in kind and pert == "interventional":
# shap 0.54 interventional output for HistGradientBoosting misses the baseline (verified error ~2 log-odds)
print("note: HistGradientBoosting -> tree_path_dependent (interventional is wrong for it in shap)")
pert = "tree_path_dependent"
if pert == "interventional":
explainer = shap.TreeExplainer(est, data=background, feature_perturbation="interventional",
model_output="raw")
else:
explainer = shap.TreeExplainer(est, feature_perturbation="tree_path_dependent")
sv, base = pick(explainer.shap_values(X, check_additivity=True), explainer.expected_value)
output = base + sv.sum(axis=1)
forest = "Forest" in kind or "DecisionTree" in kind or "ExtraTrees" in kind
space = ("probability" if forest else "raw (log-odds)") if is_clf else "prediction"
elif hasattr(est, "coef_"):
explainer = shap.LinearExplainer(est, background)
sv, base = pick(explainer.shap_values(X), explainer.expected_value)
output = base + sv.sum(axis=1)
space = "raw (log-odds)" if is_clf else "prediction"
else:
if is_clf:
k = CLASS if CLASS is not None else -1
f = lambda a: est.predict_proba(pd.DataFrame(a, columns=names))[:, k]
else:
f = lambda a: est.predict(pd.DataFrame(a, columns=names))
explainer = shap.PermutationExplainer(f, shap.maskers.Independent(background, max_samples=BACKGROUND),
seed=SEED)
e = explainer(X)
sv, base = e.values, float(np.atleast_1d(e.base_values)[0])
output = f(X.to_numpy())
space = "probability" if is_clf else "prediction"
sv = np.asarray(sv, dtype=float)
err = np.abs(base + sv.sum(axis=1) - np.asarray(output, dtype=float)).max()
assert err < 1e-3 * max(1.0, float(np.abs(output).max())), f"additivity failed: max error {{err}}"
shap_df = pd.DataFrame(sv, columns=names)
feat_df = X.copy()
if GROUP and prep is not None: # sum SHAP of derived columns back to their input column
originals = sorted(X_raw.columns, key=len, reverse=True)
def source(name):
short = name.split("__", 1)[-1]
return next((o for o in originals if short == o or short.startswith(o + "_")), short)
groups = {{}}
for n in names:
groups.setdefault(source(n), []).append(n)
shap_df = pd.DataFrame({{g: shap_df[cols].sum(axis=1) for g, cols in groups.items()}})
feat_df = X_raw[[g for g in groups if g in X_raw.columns]].copy()
shap_df = shap_df[feat_df.columns]
shap_df["base_value"] = base
shap_df["prediction"] = output
shap_df.to_csv("shap_values.csv", index=False)
feat_df.to_csv("shap_features.csv", index=False)
print(f"{{kind}}: explained {{len(X)}} rows in {{space}} space; base value {{base:.5g}}; additivity error {{err:.2e}}")
print("wrote shap_values.csv and shap_features.csv")
'''
def main():
ap = argparse.ArgumentParser(description="Generate a SHAP export 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="CSV with the rows to explain")
ap.add_argument("--target", default="", help="target column to drop if present")
ap.add_argument("--drop", nargs="*", default=[])
ap.add_argument("--rows", type=int, default=2000, help="rows to explain (sampled)")
ap.add_argument("--background", type=int, default=200, help="background sample size")
ap.add_argument("--perturbation", choices=["interventional", "tree_path_dependent"], default="interventional")
ap.add_argument("--class", dest="cls", type=int, default=None, help="class index for multiclass models")
ap.add_argument("--group", action="store_true", help="sum one-hot/derived SHAP columns per input column")
ap.add_argument("-o", "--output")
a = ap.parse_args()
if a.rows < 1 or a.background < 10:
print("error: --rows must be >= 1 and --background >= 10", file=sys.stderr)
return 2
text = TEMPLATE.format(model=a.model, data=a.data, target=a.target, drop=a.drop, rows=a.rows,
background=a.background, perturbation=a.perturbation, cls=a.cls, group=a.group,
script=os.path.basename(a.output or "explain.py"),
link_hint=" --link logit (classifiers with log-odds output)")
if a.output:
with open(a.output, "w", encoding="utf-8") as f:
f.write(text)
print(f"wrote {a.output}: {a.perturbation} SHAP, {a.rows} rows, background {a.background}"
+ (", grouped by input column" if a.group 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.