#!/usr/bin/env python3
"""Generate a learning-to-rank training script (LightGBM lambdarank and/or XGBoost rank:ndcg).
The script it writes:
- loads the CSV, sorts by query id (rankers take group sizes in row order), keeps the chosen features
- drops queries whose documents all share one label from TRAINING (no pairwise signal), keeps them for
evaluation bookkeeping
- runs GroupKFold by query (no query in both train and validation), with early stopping on NDCG@k of an
inner validation split of each training fold, and collects out-of-fold scores
- LightGBM: LGBMRanker(objective="lambdarank", label_gain = 2^label - 1, lambdarank_truncation_level),
XGBoost: XGBRanker(objective="rank:ndcg", lambdarank_pair_method="topk")
- writes oof_scores.csv (qid, label, [doc id], one column per model plus the raw features) so that
rank_eval.py compares the models with single-feature baselines on identical queries, prints NDCG@k per
model, refits each model on all data with the mean best iteration and saves it, and prints feature
importance (gain)
Needs pandas, numpy, scikit-learn, lightgbm and/or xgboost, joblib.
Usage:
python gen_ltr_script.py --data ltr.csv --qid-col qid --label-col label --features bm25 freshness clicks -o ltr.py
python gen_ltr_script.py --data train.csv --qid-col query --label-col rel --exclude doc_id --models lgbm --k 5
python gen_ltr_script.py --data ltr.csv --features bm25 clicks --models lgbm xgb --folds 5 --k 10
Exit codes: 0 written, 2 usage error.
"""
import argparse
import os
import sys
TEMPLATE = '''"""Learning-to-rank models for {data!r}. Generated by gen_ltr_script.py.
Run: python {script}
Then: python rank_eval.py oof_scores.csv --qid-col {qid} --label-col {label} --scores {score_cols} --baseline {baseline} --k {k}
"""
import joblib
import numpy as np
import pandas as pd
from sklearn.model_selection import GroupKFold, GroupShuffleSplit
DATA, QID, LABEL, FEATURES, EXCLUDE = {data!r}, {qid!r}, {label!r}, {features!r}, {exclude!r}
MODELS, K, FOLDS, SEED = {models!r}, {k}, {folds}, 0
df = pd.read_csv(DATA)
df = df.sort_values(QID, kind="stable").reset_index(drop=True) # rankers need contiguous query blocks
feats = FEATURES or [c for c in df.columns if c not in [QID, LABEL] + EXCLUDE and pd.api.types.is_numeric_dtype(df[c])]
df[LABEL] = df[LABEL].astype(int)
informative = df.groupby(QID)[LABEL].transform("nunique") > 1
print(f"{{df[QID].nunique()}} queries, {{len(df)}} documents, features: {{feats}}; "
f"{{(~informative).groupby(df[QID]).first().sum()}} single-label queries excluded from training")
def groups_of(frame):
return frame.groupby(QID, sort=False).size().to_numpy()
def make(name, n_estimators=2000):
if name == "lgbm":
from lightgbm import LGBMRanker
max_label = int(df[LABEL].max())
return LGBMRanker(objective="lambdarank", n_estimators=n_estimators, learning_rate=0.05, num_leaves=15,
min_child_samples=10, subsample=0.8, subsample_freq=1, colsample_bytree=0.8,
label_gain=[2 ** i - 1 for i in range(max_label + 1)], lambdarank_truncation_level=K,
random_state=SEED, verbose=-1)
from xgboost import XGBRanker
return XGBRanker(objective="rank:ndcg", n_estimators=n_estimators, learning_rate=0.05, max_depth=4,
subsample=0.8, colsample_bytree=0.8, lambdarank_pair_method="topk",
lambdarank_num_pair_per_sample=K, eval_metric=f"ndcg@{{K}}", early_stopping_rounds=100,
random_state=SEED)
def fit(name, model, tr, va):
if name == "lgbm":
import lightgbm as lgb
model.fit(tr[feats], tr[LABEL], group=groups_of(tr), eval_set=[(va[feats], va[LABEL])],
eval_group=[groups_of(va)], eval_at=[K], callbacks=[lgb.early_stopping(100, verbose=False)])
return model.best_iteration_ or model.n_estimators
model.fit(tr[feats], tr[LABEL], qid=tr[QID].astype("category").cat.codes, eval_set=[(va[feats], va[LABEL])],
eval_qid=[va[QID].astype("category").cat.codes], verbose=False)
return (model.best_iteration or model.n_estimators - 1) + 1
oof = df[[QID, LABEL] + [c for c in df.columns if c not in [QID, LABEL] and c not in feats] + feats].copy()
best_iters = {{m: [] for m in MODELS}}
gkf = GroupKFold(n_splits=FOLDS)
for fold, (tr_idx, te_idx) in enumerate(gkf.split(df, groups=df[QID])):
train_part = df.iloc[tr_idx]
train_part = train_part[informative.iloc[tr_idx].to_numpy()]
inner = GroupShuffleSplit(n_splits=1, test_size=0.2, random_state=SEED)
a_idx, v_idx = next(inner.split(train_part, groups=train_part[QID]))
tr = train_part.iloc[a_idx].sort_values(QID, kind="stable")
va = train_part.iloc[v_idx].sort_values(QID, kind="stable")
for m in MODELS:
model = make(m)
best_iters[m].append(fit(m, model, tr, va))
oof.loc[df.index[te_idx], m] = model.predict(df.iloc[te_idx][feats])
def ndcg_at(frame, col):
vals = []
for _, g in frame.groupby(QID):
# ties broken pessimistically (lower label first), as in rank_eval.py: tied scores must not look good
lab = g.sort_values([col, LABEL], ascending=[False, True], kind="stable")[LABEL].to_numpy()[:K]
ideal = np.sort(g[LABEL].to_numpy())[::-1][:K]
disc = 1 / np.log2(np.arange(2, len(lab) + 2))
idcg = ((2.0 ** ideal - 1) * disc[:len(ideal)]).sum()
if idcg > 0:
vals.append(((2.0 ** lab - 1) * disc).sum() / idcg)
return float(np.mean(vals))
oof.to_csv("oof_scores.csv", index=False)
for m in MODELS:
print(f"{{m}}: out-of-fold NDCG@{{K}} {{ndcg_at(oof, m):.4f}} (best iterations {{best_iters[m]}})")
for f in feats:
print(f"single feature {{f}}: NDCG@{{K}} {{ndcg_at(oof, f):.4f}}")
full = df[informative.to_numpy()]
for m in MODELS:
n_iter = max(10, int(np.mean(best_iters[m]) * 1.1))
model = make(m, n_estimators=n_iter)
if m == "lgbm":
model.fit(full[feats], full[LABEL], group=groups_of(full))
imp = dict(zip(feats, model.booster_.feature_importance(importance_type="gain")))
else:
model.set_params(early_stopping_rounds=None)
model.fit(full[feats], full[LABEL], qid=full[QID].astype("category").cat.codes)
score = model.get_booster().get_score(importance_type="total_gain")
imp = {{f: score.get(f, 0.0) for f in feats}}
tot = sum(imp.values()) or 1.0
print(f"{{m}} gain importance: " + ", ".join(f"{{f}} {{v / tot:.0%}}" for f, v in sorted(imp.items(), key=lambda kv: -kv[1])))
joblib.dump(model, f"ranker_{{m}}.joblib")
print("wrote oof_scores.csv and ranker_<model>.joblib")
'''
def main():
ap = argparse.ArgumentParser(description="Generate a learning-to-rank training script.", epilog=__doc__,
formatter_class=argparse.RawDescriptionHelpFormatter)
ap.add_argument("--data", required=True)
ap.add_argument("--qid-col", default="qid")
ap.add_argument("--label-col", default="label")
ap.add_argument("--features", nargs="*", default=[])
ap.add_argument("--exclude", nargs="*", default=[])
ap.add_argument("--models", nargs="+", choices=["lgbm", "xgb"], default=["lgbm"])
ap.add_argument("--k", type=int, default=10)
ap.add_argument("--folds", type=int, default=5)
ap.add_argument("-o", "--output")
a = ap.parse_args()
if a.k < 1 or a.folds < 2:
print("error: --k must be >= 1 and --folds >= 2", file=sys.stderr)
return 2
baseline = a.features[0] if a.features else "<feature>"
text = TEMPLATE.format(data=a.data, qid=a.qid_col, label=a.label_col, features=a.features, exclude=a.exclude,
models=a.models, k=a.k, folds=a.folds, score_cols=" ".join(a.models), baseline=baseline,
script=os.path.basename(a.output or "ltr.py"))
if a.output:
with open(a.output, "w", encoding="utf-8") as f:
f.write(text)
print(f"wrote {a.output}: {', '.join(a.models)}, NDCG@{a.k}, {a.folds} grouped folds", 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.