#!/usr/bin/env python3
"""Generate a fraud model training script: point-in-time features, out-of-time split, LightGBM + anomaly score.
The script it writes:
- sorts transactions by time and builds per-entity velocity features with pandas rolling windows that
exclude the current transaction (closed="left"): count and amount over --windows, time since the
previous transaction, amount relative to the entity's past mean, first-time category/country flags
- splits OUT OF TIME: train on the earliest part, leave a --gap-days gap (labels such as chargebacks mature
weeks after the transaction; the most recent training labels are incomplete), test on the latest part
- adds an IsolationForest anomaly score fitted on the training period only (unsupervised signal for new
fraud patterns)
- trains LightGBM (no resampling; scale_pos_weight optional) with early stopping on average precision of
a time-ordered validation tail of the training period
- writes test_scored.csv (label, score, amount, timestamp, id) for fraud_eval.py and prints test ROC AUC,
average precision and gain importance
Needs pandas, numpy, scikit-learn, lightgbm, joblib.
Usage:
python gen_fraud_model_script.py --data transactions.csv --entity card_id --label-col is_fraud -o fraud_model.py
python gen_fraud_model_script.py --data tx.csv --entity account_id --categorical merchant_category country \\
--test-frac 0.3 --gap-days 14 --windows 1h 24h 7d 30d
Exit codes: 0 written, 2 usage error.
"""
import argparse
import os
import sys
TEMPLATE = '''"""Fraud model for {data!r}. Generated by gen_fraud_model_script.py.
Run: python {script}
Then: python fraud_eval.py test_scored.csv --label-col {label} --score-col score --amount-col {amount} --time-col {time}
"""
import joblib
import lightgbm as lgb
import numpy as np
import pandas as pd
from sklearn.ensemble import IsolationForest
from sklearn.metrics import average_precision_score, roc_auc_score
DATA, ENTITY, TIME, AMOUNT, LABEL, ID = {data!r}, {entity!r}, {time!r}, {amount!r}, {label!r}, {id_col!r}
CATEGORICAL, WINDOWS, TEST_FRAC, GAP_DAYS, POS_WEIGHT, SEED = {categorical!r}, {windows!r}, {test_frac}, {gap_days}, {pos_weight}, 0
df = pd.read_csv(DATA, parse_dates=[TIME]).sort_values(TIME, kind="stable").reset_index(drop=True)
df["_row"] = np.arange(len(df))
# ---- point-in-time features: every window excludes the current transaction (closed="left")
g = df.set_index(TIME).groupby(ENTITY)[AMOUNT]
for w in WINDOWS:
roll = g.rolling(w, closed="left").agg(["count", "sum"]).reset_index()
roll["_row"] = df.sort_values([ENTITY, TIME], kind="stable")["_row"].to_numpy()
roll = roll.set_index("_row").sort_index()
df[f"cnt_{{w}}"] = roll["count"].fillna(0).to_numpy()
df[f"sum_{{w}}"] = roll["sum"].fillna(0).to_numpy()
prev_t = df.groupby(ENTITY)[TIME].shift(1)
df["secs_since_prev"] = (df[TIME] - prev_t).dt.total_seconds()
past_mean = df.groupby(ENTITY)[AMOUNT].transform(lambda s: s.shift(1).expanding().mean())
df["amt_vs_mean"] = df[AMOUNT] / past_mean
df["n_prev"] = df.groupby(ENTITY).cumcount()
for c in CATEGORICAL:
first = ~df.duplicated([ENTITY, c])
df[f"new_{{c}}"] = first.astype(int)
df[c] = df[c].astype("category")
features = ([AMOUNT] + [f"cnt_{{w}}" for w in WINDOWS] + [f"sum_{{w}}" for w in WINDOWS]
+ ["secs_since_prev", "amt_vs_mean", "n_prev"] + [f"new_{{c}}" for c in CATEGORICAL] + CATEGORICAL)
df["hour"] = df[TIME].dt.hour
features.append("hour")
# ---- out-of-time split with a label-maturity gap
t_test = df[TIME].quantile(1 - TEST_FRAC)
t_train_end = t_test - pd.to_timedelta(GAP_DAYS, unit="D")
train = df[df[TIME] < t_train_end].copy()
test = df[df[TIME] >= t_test].copy()
print(f"train {{len(train)}} rows to {{t_train_end:%Y-%m-%d}} ({{train[LABEL].mean():.2%}} fraud), gap {{GAP_DAYS}} days, "
f"test {{len(test)}} rows from {{t_test:%Y-%m-%d}} ({{test[LABEL].mean():.2%}} fraud)")
if train[LABEL].sum() < 10 or test[LABEL].sum() < 5:
print("warning: very few frauds in train or test; metrics will be noisy")
# ---- unsupervised anomaly score (fit on the training period only)
num_feats = [f for f in features if f not in CATEGORICAL]
iso = IsolationForest(n_estimators=300, random_state=SEED).fit(train[num_feats].fillna(-1))
for part in (train, test):
part.loc[:, "iso_score"] = -iso.score_samples(part[num_feats].fillna(-1))
features.append("iso_score")
# ---- LightGBM with a time-ordered validation tail for early stopping
cut = int(len(train) * 0.8)
tr, va = train.iloc[:cut], train.iloc[cut:]
model = lgb.LGBMClassifier(n_estimators=3000, learning_rate=0.03, num_leaves=15, min_child_samples=20,
subsample=0.8, subsample_freq=1, colsample_bytree=0.8, reg_lambda=1.0,
scale_pos_weight=POS_WEIGHT, random_state=SEED, verbose=-1)
model.fit(tr[features], tr[LABEL], eval_set=[(va[features], va[LABEL])], eval_metric="average_precision",
callbacks=[lgb.early_stopping(200, verbose=False)])
best = model.best_iteration_ or 200
final = lgb.LGBMClassifier(**{{**model.get_params(), "n_estimators": max(50, int(best * 1.1))}})
final.fit(train[features], train[LABEL])
score = final.predict_proba(test[features])[:, 1]
print(f"test ROC AUC {{roc_auc_score(test[LABEL], score):.4f}} average precision {{average_precision_score(test[LABEL], score):.4f}} "
f"(random {{test[LABEL].mean():.4f}}); IsolationForest alone AP {{average_precision_score(test[LABEL], test['iso_score']):.4f}}")
imp = dict(zip(features, final.booster_.feature_importance(importance_type="gain")))
tot = sum(imp.values()) or 1.0
print("gain importance: " + ", ".join(f"{{k}} {{v / tot:.0%}}" for k, v in sorted(imp.items(), key=lambda kv: -kv[1])[:8]))
out = test[[c for c in [ID, ENTITY, TIME, AMOUNT, LABEL] if c]].copy()
out["score"] = score
out.to_csv("test_scored.csv", index=False)
joblib.dump({{"model": final, "iso": iso, "features": features}}, "fraud_model.joblib")
print("wrote test_scored.csv and fraud_model.joblib")
'''
def main():
ap = argparse.ArgumentParser(description="Generate a fraud model training script.", epilog=__doc__,
formatter_class=argparse.RawDescriptionHelpFormatter)
ap.add_argument("--data", required=True)
ap.add_argument("--entity", required=True)
ap.add_argument("--time-col", default="timestamp")
ap.add_argument("--amount-col", default="amount")
ap.add_argument("--label-col", default="is_fraud")
ap.add_argument("--id-col", default="txn_id")
ap.add_argument("--categorical", nargs="*", default=[])
ap.add_argument("--windows", nargs="+", default=["1h", "24h", "7d"])
ap.add_argument("--test-frac", type=float, default=0.3)
ap.add_argument("--gap-days", type=float, default=7)
ap.add_argument("--pos-weight", type=float, default=1.0, help="LightGBM scale_pos_weight (1 = none)")
ap.add_argument("-o", "--output")
a = ap.parse_args()
if not 0.1 <= a.test_frac <= 0.5 or a.gap_days < 0:
print("error: --test-frac must be in [0.1, 0.5] and --gap-days >= 0", file=sys.stderr)
return 2
text = TEMPLATE.format(data=a.data, entity=a.entity, time=a.time_col, amount=a.amount_col, label=a.label_col,
id_col=a.id_col, categorical=a.categorical, windows=a.windows, test_frac=a.test_frac,
gap_days=a.gap_days, pos_weight=a.pos_weight,
script=os.path.basename(a.output or "fraud_model.py"))
if a.output:
with open(a.output, "w", encoding="utf-8") as f:
f.write(text)
print(f"wrote {a.output}: entity {a.entity}, windows {' '.join(a.windows)}, out-of-time test {a.test_frac:.0%} "
f"after a {a.gap_days:g}-day gap", 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.