#!/usr/bin/env python3
"""Build (or audit existing) cross-validation folds and check them before any model is trained.
Either generate folds with --strategy kfold | stratified | group | stratified-group | time, or audit a
column of existing fold ids with --fold-col. For every fold it reports the train/test sizes, the target
rate (classification) or mean (regression) in the test fold, and three leakage checks:
group overlap number of --group-col values present in both train and test (must be 0 for grouped CV)
time order latest training time vs earliest test time (forward folds need train_max < test_min,
and a gap of at least --label-window days when labels look ahead)
coverage classes missing from a test fold, very small test folds, unequal fold sizes
Generated folds follow scikit-learn's logic (shuffled stratification, greedy group balancing,
forward-chaining blocks with a gap), so the preview predicts what scikit-learn will do.
Usage:
python fold_preview.py data.csv --target churned --strategy stratified --folds 5
python fold_preview.py txns.csv --target is_fraud --strategy stratified-group --group-col card_id
python fold_preview.py snaps.csv --target churned --strategy time --time-col score_date --label-window 30
python fold_preview.py data.csv --target y --fold-col fold --group-col customer_id --time-col ts
Exit codes: 0 folds look sound, 1 a leakage or coverage problem was found, 2 usage error.
"""
import argparse
import csv
import datetime as dt
import json
import math
import random
import sys
from collections import Counter, defaultdict
MISSING = {"", "na", "n/a", "nan", "null", "none", "?"}
def parse_time(v):
v = (v or "").strip()
try:
return dt.datetime.fromisoformat(v.replace("Z", "+00:00")).replace(tzinfo=None)
except ValueError:
try:
return dt.datetime.fromtimestamp(float(v))
except (ValueError, OverflowError, OSError):
return None
def make_folds(rows, a, y_cls):
n = len(rows)
rng = random.Random(a.seed)
k = a.folds
fold = [0] * n
if a.strategy == "kfold":
idx = list(range(n))
rng.shuffle(idx)
for pos, i in enumerate(idx):
fold[i] = pos * k // n
elif a.strategy == "stratified":
by = defaultdict(list)
for i, c in enumerate(y_cls):
by[c].append(i)
for c, idx in by.items():
rng.shuffle(idx)
for pos, i in enumerate(idx):
fold[i] = pos % k
elif a.strategy in ("group", "stratified-group"):
members = defaultdict(list)
for i, r in enumerate(rows):
members[r[a.group_col]].append(i)
groups = list(members)
rng.shuffle(groups)
if a.strategy == "group":
groups.sort(key=lambda g: -len(members[g])) # greedy: largest group to the smallest fold
sizes = [0] * k
for g in groups:
f = min(range(k), key=lambda j: sizes[j])
for i in members[g]:
fold[i] = f
sizes[f] += len(members[g])
else:
classes = sorted(set(y_cls))
target = Counter(y_cls)
counts = [Counter() for _ in range(k)]
groups.sort(key=lambda g: -max(Counter(y_cls[i] for i in members[g]).values()))
for g in groups:
gc = Counter(y_cls[i] for i in members[g])
def cost(j):
trial = [counts[j][c] + gc[c] if jj == j else counts[jj][c] for jj in range(k) for c in classes]
devs = []
for c_i, c in enumerate(classes):
per = [trial[jj * len(classes) + c_i] / target[c] for jj in range(k)]
m = sum(per) / k
devs.append(math.sqrt(sum((p - m) ** 2 for p in per) / k))
return sum(devs) / len(devs), sum(counts[j].values())
f = min(range(k), key=cost)
counts[f].update(gc)
for i in members[g]:
fold[i] = f
return fold
def main():
ap = argparse.ArgumentParser(description="Preview or audit CV folds for leakage and balance.", epilog=__doc__,
formatter_class=argparse.RawDescriptionHelpFormatter)
ap.add_argument("csv")
ap.add_argument("--target", required=True)
ap.add_argument("--strategy", choices=["kfold", "stratified", "group", "stratified-group", "time"])
ap.add_argument("--fold-col", help="audit this existing fold-id column instead of generating folds")
ap.add_argument("--folds", type=int, default=5)
ap.add_argument("--group-col")
ap.add_argument("--time-col")
ap.add_argument("--label-window", type=float, default=0, help="days of label look-ahead (time folds)")
ap.add_argument("--gap", type=int, default=0, help="rows dropped between train and test (time strategy)")
ap.add_argument("--seed", type=int, default=0)
ap.add_argument("--json", action="store_true")
a = ap.parse_args()
if not (a.strategy or a.fold_col):
print("error: give --strategy or --fold-col", file=sys.stderr)
return 2
try:
with open(a.csv, newline="", encoding="utf-8-sig") as fh:
rd = csv.DictReader(fh)
header = rd.fieldnames or []
rows = [r for r in rd if (r.get(a.target) or "").strip().lower() not in MISSING]
except OSError as e:
print(f"error: {e}", file=sys.stderr)
return 2
for c in (a.target, a.fold_col, a.group_col, a.time_col):
if c and c not in header:
print(f"error: column {c!r} not found", file=sys.stderr)
return 2
if a.strategy in ("group", "stratified-group") and not a.group_col:
print("error: group strategies need --group-col", file=sys.stderr)
return 2
if a.strategy == "time" and not a.time_col:
print("error: --strategy time needs --time-col", file=sys.stderr)
return 2
yv = [r[a.target].strip() for r in rows]
classes = Counter(yv)
classification = len(classes) <= 20
times = [parse_time(r[a.time_col]) for r in rows] if a.time_col else None
n = len(rows)
splits = [] # list of (train_idx, test_idx)
if a.fold_col:
ids = sorted({r[a.fold_col] for r in rows}, key=lambda v: (len(v), v))
for f in ids:
te = [i for i, r in enumerate(rows) if r[a.fold_col] == f]
tr = [i for i, r in enumerate(rows) if r[a.fold_col] != f]
if times:
tr = [i for i in tr if times[i] and times[i] < min(times[j] for j in te if times[j])] \
if a.strategy == "time" else tr
splits.append((tr, te))
elif a.strategy == "time":
order = sorted(range(n), key=lambda i: times[i] or dt.datetime.min)
test_size = n // (a.folds + 1)
for f in range(a.folds):
te_start = n - (a.folds - f) * test_size
te = order[te_start:te_start + test_size]
tr = order[:max(0, te_start - a.gap)]
splits.append((tr, te))
else:
fold = make_folds(rows, a, yv)
for f in range(a.folds):
splits.append(([i for i in range(n) if fold[i] != f], [i for i in range(n) if fold[i] == f]))
problems, report = [], []
for f, (tr, te) in enumerate(splits):
e = {"fold": f, "train": len(tr), "test": len(te)}
if classification:
tc = Counter(yv[i] for i in te)
minority = min(classes, key=classes.get)
e["test_minority_rate"] = tc.get(minority, 0) / max(1, len(te))
miss = [c for c in classes if tc.get(c, 0) == 0]
if miss:
problems.append(f"fold {f}: classes {miss} absent from the test fold")
else:
vals = [float(yv[i]) for i in te]
e["test_target_mean"] = sum(vals) / max(1, len(vals))
if a.group_col:
gtr = {rows[i][a.group_col] for i in tr}
overlap = sum(1 for g in {rows[i][a.group_col] for i in te} if g in gtr)
e["group_overlap"] = overlap
if overlap and a.strategy in ("group", "stratified-group", None):
problems.append(f"fold {f}: {overlap} {a.group_col} value(s) in both train and test")
elif overlap and a.strategy in ("kfold", "stratified"):
problems.append(f"fold {f}: {overlap} {a.group_col} value(s) shared between train and test "
"(random folds with repeated entities)")
if times and tr and te:
tmax = max(times[i] for i in tr if times[i])
tmin = min(times[i] for i in te if times[i])
gap_days = (tmin - tmax).total_seconds() / 86400
e["train_max"], e["test_min"], e["gap_days"] = str(tmax), str(tmin), round(gap_days, 2)
if a.strategy == "time" or a.fold_col:
if gap_days < 0:
problems.append(f"fold {f}: training data ends after the test fold starts ({gap_days:.1f} days)")
elif gap_days < a.label_window:
problems.append(f"fold {f}: gap {gap_days:.1f} days < label window {a.label_window:g}: training "
"labels overlap the test period")
if len(te) < 0.5 * n / max(1, len(splits)) and a.strategy != "time":
problems.append(f"fold {f}: test fold has only {len(te)} rows")
report.append(e)
out = {"rows": n, "strategy": a.strategy or f"existing column {a.fold_col}", "folds": report, "problems": problems}
if a.json:
print(json.dumps(out, indent=2))
return 1 if problems else 0
print(f"{n} rows, {out['strategy']}, {len(splits)} folds")
for e in report:
bits = [f"train {e['train']:>6}", f"test {e['test']:>5}"]
if "test_minority_rate" in e:
bits.append(f"minority {e['test_minority_rate']:.3f}")
if "test_target_mean" in e:
bits.append(f"mean {e['test_target_mean']:.4g}")
if "group_overlap" in e:
bits.append(f"group overlap {e['group_overlap']}")
if "gap_days" in e:
bits.append(f"train<= {e['train_max'][:10]} test>= {e['test_min'][:10]} gap {e['gap_days']:g} d")
print(f" fold {e['fold']}: " + " ".join(bits))
for p in problems:
print(f"PROBLEM: {p}")
if not problems:
print("folds look sound: no group overlap, correct time order, every class present")
return 1 if problems else 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.