#!/usr/bin/env python3
"""Generate a survival modelling script: Cox, Weibull AFT, random survival forest and gradient boosting,
compared with survival-appropriate metrics on a held-out set.
The script it writes:
- loads the CSV, one-hot encodes categorical covariates (drop first), splits train/test stratified on the
event indicator (or by --time-col order for out-of-time evaluation)
- fits lifelines CoxPHFitter (penalizer chosen from a small grid by 5-fold CV concordance), lifelines
WeibullAFTFitter, and scikit-survival RandomSurvivalForest and GradientBoostingSurvivalAnalysis
- evaluates on the test set: Harrell's C, Uno's C (IPCW, truncated at tau), time-dependent cumulative/
dynamic AUC at --times, integrated Brier score over the --times range (lower is better, 0.25 = useless
for S near 0.5), and calibration at the middle time: predicted vs Kaplan-Meier observed survival by
risk quintile
- writes survival_model_comparison.csv, test_survival_predictions.csv (S(t) per test row and time) and
the best model by integrated Brier score with joblib
Needs lifelines, scikit-survival, scikit-learn, pandas, numpy, joblib.
Usage:
python gen_survival_script.py --data survival.csv --duration-col duration_days --event-col event --covariates plan age -o surv.py
python gen_survival_script.py --data data.csv --duration-col t --event-col died --covariates age stage sex --times 90 180 365
python gen_survival_script.py --data data.csv --duration-col t --event-col churn --covariates plan --time-col signup_date
Exit codes: 0 written, 2 usage error.
"""
import argparse
import os
import sys
TEMPLATE = '''"""Survival models for {duration!r} / {event!r}. Generated by gen_survival_script.py.
Run: python {script}
"""
import joblib
import numpy as np
import pandas as pd
from lifelines import CoxPHFitter, KaplanMeierFitter, WeibullAFTFitter
from lifelines.utils import k_fold_cross_validation
from sklearn.model_selection import train_test_split
from sksurv.ensemble import GradientBoostingSurvivalAnalysis, RandomSurvivalForest
from sksurv.metrics import (concordance_index_censored, concordance_index_ipcw, cumulative_dynamic_auc,
integrated_brier_score)
from sksurv.util import Surv
DATA, DURATION, EVENT, COVARIATES, TIME_COL = {data!r}, {duration!r}, {event!r}, {covariates!r}, {time_col!r}
TIMES, TEST_SIZE, SEED = {times!r}, {test_size}, 0
df = pd.read_csv(DATA)
df = df.dropna(subset=[DURATION, EVENT] + COVARIATES)
df = df[df[DURATION] > 0]
X = pd.get_dummies(df[COVARIATES], drop_first=True, dtype=float)
data = pd.concat([X, df[[DURATION, EVENT]]], axis=1)
if TIME_COL:
order = df[TIME_COL].sort_values().index
cut = int(len(order) * (1 - TEST_SIZE))
train, test = data.loc[order[:cut]], data.loc[order[cut:]]
else:
train, test = train_test_split(data, test_size=TEST_SIZE, stratify=data[EVENT], random_state=SEED)
feats = list(X.columns)
y_tr = Surv.from_arrays(train[EVENT].astype(bool), train[DURATION])
y_te = Surv.from_arrays(test[EVENT].astype(bool), test[DURATION])
if TIMES:
times = np.array(TIMES, dtype=float)
else: # inside the test follow-up, where IPCW metrics are defined
times = np.quantile(test.loc[test[EVENT] == 1, DURATION], [0.2, 0.4, 0.6, 0.8])
times = times[(times > test[DURATION].min()) & (times < test[DURATION].max()) & (times < train[DURATION].max())]
tau = float(times.max())
def cox_factory():
best = None
for pen in (0.0, 0.01, 0.1, 1.0):
scores = k_fold_cross_validation(CoxPHFitter(penalizer=pen), train, DURATION, EVENT, k=5,
scoring_method="concordance_index", seed=SEED)
if best is None or np.mean(scores) > best[0]:
best = (np.mean(scores), pen)
print(f"Cox penalizer {{best[1]}} (CV concordance {{best[0]:.3f}})")
return CoxPHFitter(penalizer=best[1]).fit(train, DURATION, EVENT)
models = {{}}
models["cox"] = cox_factory()
models["weibull_aft"] = WeibullAFTFitter(penalizer=0.01).fit(train, DURATION, EVENT)
models["random_survival_forest"] = RandomSurvivalForest(n_estimators=300, min_samples_leaf=10, max_features="sqrt",
random_state=SEED, n_jobs=-1).fit(train[feats], y_tr)
models["gradient_boosting"] = GradientBoostingSurvivalAnalysis(n_estimators=200, learning_rate=0.05, max_depth=2,
subsample=0.8, random_state=SEED).fit(train[feats], y_tr)
def surv_matrix(name, m, frame):
"""Rows: test subjects, columns: times -> S(t)."""
if name in ("cox", "weibull_aft"):
sf = m.predict_survival_function(frame[feats], times=times)
return sf.T.to_numpy()
fns = m.predict_survival_function(frame[feats])
return np.vstack([fn(times) for fn in fns])
def risk_score(name, m, frame):
if name == "cox":
return m.predict_partial_hazard(frame[feats]).to_numpy()
if name == "weibull_aft":
return -m.predict_median(frame[feats]).to_numpy() # shorter median = higher risk
return m.predict(frame[feats])
rows, preds = [], []
mid = times[len(times) // 2]
for name, m in models.items():
S = surv_matrix(name, m, test)
risk = risk_score(name, m, test)
harrell = concordance_index_censored(test[EVENT].astype(bool), test[DURATION], risk)[0]
uno = concordance_index_ipcw(y_tr, y_te, risk, tau=tau)[0]
auc_t, mean_auc = cumulative_dynamic_auc(y_tr, y_te, risk, times)
ibs = integrated_brier_score(y_tr, y_te, S, times) if len(times) > 1 else float("nan")
rows.append({{"model": name, "harrell_c": harrell, "uno_c": uno, "mean_auc": mean_auc, "ibs": ibs,
**{{f"auc_t{{t:g}}": a for t, a in zip(times, auc_t)}}}})
k = int(np.argmin(np.abs(times - mid)))
q = pd.qcut(pd.Series(S[:, k]).rank(method="first"), 5, labels=False)
cal = []
for g in range(5):
idx = np.where(q.to_numpy() == g)[0]
km = KaplanMeierFitter().fit(test[DURATION].iloc[idx], test[EVENT].iloc[idx])
cal.append((float(S[idx, k].mean()), float(km.survival_function_at_times(times[k]).iloc[0])))
rows[-1]["calibration_max_gap"] = max(abs(p - o) for p, o in cal)
for i in range(len(test)):
preds.append({{"model": name, "row": int(test.index[i]), **{{f"S_{{t:g}}": S[i, j] for j, t in enumerate(times)}}}})
table = pd.DataFrame(rows).sort_values("ibs")
table.to_csv("survival_model_comparison.csv", index=False)
pd.DataFrame(preds).to_csv("test_survival_predictions.csv", index=False)
print(f"test rows: {{len(test)}}, events: {{int(test[EVENT].sum())}}, times: {{', '.join(f'{{t:g}}' for t in times)}}")
print(table.round(4).to_string(index=False))
best = table.iloc[0]["model"]
joblib.dump(models[best], f"best_survival_model_{{best}}.joblib")
print(f"best by integrated Brier score: {{best}} -> best_survival_model_{{best}}.joblib")
cox = models["cox"]
try:
from lifelines.statistics import proportional_hazard_test
ph = proportional_hazard_test(cox, train, time_transform="rank").summary
bad = ph[ph["p"] < 0.05]
if len(bad):
print("Cox PH assumption violated for:", ", ".join(map(str, bad.index)))
except Exception as exc: # diagnostics only; never block the comparison
print(f"PH test skipped: {{exc}}")
'''
def main():
ap = argparse.ArgumentParser(description="Generate a survival model comparison script.", epilog=__doc__,
formatter_class=argparse.RawDescriptionHelpFormatter)
ap.add_argument("--data", required=True)
ap.add_argument("--duration-col", required=True)
ap.add_argument("--event-col", required=True)
ap.add_argument("--covariates", nargs="+", required=True)
ap.add_argument("--times", type=float, nargs="*", help="evaluation times (default: event-time quantiles)")
ap.add_argument("--time-col", help="order column for an out-of-time test split")
ap.add_argument("--test-size", type=float, default=0.3)
ap.add_argument("-o", "--output")
a = ap.parse_args()
if not 0.1 <= a.test_size <= 0.5:
print("error: --test-size must be between 0.1 and 0.5", file=sys.stderr)
return 2
text = TEMPLATE.format(data=a.data, duration=a.duration_col, event=a.event_col, covariates=a.covariates,
time_col=a.time_col, times=a.times or [], test_size=a.test_size,
script=os.path.basename(a.output or "survival_models.py"))
if a.output:
with open(a.output, "w", encoding="utf-8") as f:
f.write(text)
print(f"wrote {a.output}: Cox, Weibull AFT, RSF, GBSA on {', '.join(a.covariates)}", 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.