#!/usr/bin/env python3
"""Score the baselines every model must beat, on a holdout, with no ML libraries.
Classification (binary target):
majority always predict the most frequent class (accuracy floor, AUC 0.5)
prior predict the training positive rate as the probability (log loss / Brier floor)
one_r best single feature, chosen on an inner split of the training rows: smoothed
positive rate per level (categorical) or per 10 quantile bins (numeric)
rule your current business rule, e.g. --rule "support_calls >= 3 or plan == basic"
Regression (numeric target):
mean, median constant predictions
one_r best single feature: mean target per level or per 10 quantile bins
rule not used for regression
Split: last --test-frac of rows by --time-col (time-ordered holdout, the honest default for
anything deployed forward in time), or a seeded random split without --time-col.
Usage:
python baseline_scores.py data.csv --target churned --time-col signup_date --drop customer_id
python baseline_scores.py data.csv --target churned --rule "support_calls >= 3 or last_login_days > 30"
python baseline_scores.py data.csv --target price --drop house_id --json
Exit codes: 0 scored, 2 usage/read error or unusable target.
"""
import argparse
import csv
import json
import math
import random
import re
import sys
from collections import Counter, defaultdict
MISSING = {"", "na", "n/a", "nan", "null", "none", "?"}
RULE_TERM = re.compile(r"^\s*([A-Za-z_][\w.]*)\s*(>=|<=|==|!=|>|<)\s*(.+?)\s*$")
def num(v):
try:
x = float(v)
return x if math.isfinite(x) else None
except (TypeError, ValueError):
return None
def missing(v):
return v is None or v.strip().lower() in MISSING
# ---------------------------------------------------------------- metrics
def roc_auc(s, y):
pos = sum(y)
neg = len(y) - pos
if not pos or not neg:
return float("nan")
order = sorted(range(len(s)), key=s.__getitem__)
rank = [0.0] * len(s)
i = 0
while i < len(order):
j = i
while j + 1 < len(order) and s[order[j + 1]] == s[order[i]]:
j += 1
for k in range(i, j + 1):
rank[order[k]] = (i + j) / 2 + 1
i = j + 1
return (sum(r for r, t in zip(rank, y) if t) - pos * (pos + 1) / 2) / (pos * neg)
def avg_precision(s, y):
"""Average precision with tied scores treated as one threshold (a constant score gives the prevalence)."""
pos = sum(y)
if not pos:
return float("nan")
by_score = defaultdict(lambda: [0, 0])
for v, t in zip(s, y):
by_score[v][0] += t
by_score[v][1] += 1
tp = seen = 0
total = 0.0
for v in sorted(by_score, reverse=True):
hits, n = by_score[v]
tp += hits
seen += n
total += hits / pos * (tp / seen)
return total
def cls_metrics(prob, y, thr):
"""accuracy at 0.5; balanced accuracy at thr (the training positive rate)."""
pos, neg = sum(y), len(y) - sum(y)
acc = sum(1 for p, t in zip(prob, y) if (p >= 0.5) == bool(t)) / len(y)
tp = sum(1 for p, t in zip(prob, y) if p >= thr and t)
tn = sum(1 for p, t in zip(prob, y) if p < thr and not t)
eps = 1e-15
return {
"accuracy": acc,
"balanced_acc": ((tp / pos if pos else 0) + (tn / neg if neg else 0)) / 2,
"roc_auc": roc_auc(prob, y),
"pr_auc": avg_precision(prob, y),
"log_loss": -sum(t * math.log(min(1 - eps, max(eps, p))) + (1 - t) * math.log(min(1 - eps, max(eps, 1 - p)))
for p, t in zip(prob, y)) / len(y),
"brier": sum((p - t) ** 2 for p, t in zip(prob, y)) / len(y),
}
def reg_metrics(pred, y):
n = len(y)
mean = sum(y) / n
sse = sum((p - t) ** 2 for p, t in zip(pred, y))
sst = sum((t - mean) ** 2 for t in y)
abs_err = sum(abs(p - t) for p, t in zip(pred, y))
return {"mae": abs_err / n, "rmse": math.sqrt(sse / n), "r2": 1 - sse / sst if sst else float("nan"),
"wape": abs_err / sum(abs(t) for t in y) if any(y) else float("nan")}
# ---------------------------------------------------------------- one-feature learners
def fit_one_feature(train, col, y_tr, label):
"""Smoothed mean target per level (categorical) or per 10 quantile bins (numeric) of one column."""
vals = [r.get(col, "") for r in train]
nums = [num(v) for v in vals]
prior = sum(y_tr) / len(y_tr)
if sum(x is not None for x in nums) / len(nums) > 0.95:
xs = sorted(x for x in nums if x is not None)
edges = sorted({xs[int(len(xs) * q / 10)] for q in range(1, 10)})
def key(r):
x = num(r.get(col, ""))
return None if x is None else sum(1 for e in edges if x >= e)
desc = f"{label} by {len(edges) + 1} bins of {col}"
else:
def key(r):
return r.get(col, "")
desc = f"{label} by {col} ({len(set(vals))} levels)"
sums = defaultdict(lambda: [0.0, 0])
for r, t in zip(train, y_tr):
b = sums[key(r)]
b[0] += t
b[1] += 1
means = {k: (sm + 5 * prior) / (n + 5) for k, (sm, n) in sums.items()}
return {"desc": desc, "score": lambda r: means.get(key(r), prior)}
def pick_one_feature(train, feats, y_tr, binary):
"""Choose the column on an inner split of the training rows (never on the test set), refit on all of train."""
cut = int(len(train) * 0.7)
fit, val, y_fit, y_val = train[:cut], train[cut:], y_tr[:cut], y_tr[cut:]
best = None
for c in feats:
m = fit_one_feature(fit, c, y_fit, "rate" if binary else "mean")
pr = [m["score"](r) for r in val]
q = roc_auc(pr, y_val) if binary else -reg_metrics(pr, y_val)["mae"]
if q == q and (best is None or q > best[0]):
best = (q, c)
if best is None:
return None
return fit_one_feature(train, best[1], y_tr, "positive rate" if binary else "mean")
# ---------------------------------------------------------------- rule
def compile_rule(text, header):
"""'a >= 3 and b == x or c < 2' -> predicate; 'and' binds tighter than 'or'; no parentheses."""
ors = []
for disj in re.split(r"\s+or\s+", text.strip()):
ands = []
for term in re.split(r"\s+and\s+", disj):
m = RULE_TERM.match(term)
if not m:
raise ValueError(f"cannot parse rule term {term!r}; use 'column op value'")
col, op, val = m.group(1), m.group(2), m.group(3).strip("'\"")
if col not in header:
raise ValueError(f"rule column {col!r} not in data")
ands.append((col, op, val))
ors.append(ands)
def holds(r, col, op, val):
a, b = num(r.get(col, "")), num(val)
if a is not None and b is not None:
return {">=": a >= b, "<=": a <= b, ">": a > b, "<": a < b, "==": a == b, "!=": a != b}[op]
s = (r.get(col) or "").strip()
return {"==": s == val, "!=": s != val}.get(op, False)
return lambda r: any(all(holds(r, *t) for t in ands) for ands in ors)
def main():
ap = argparse.ArgumentParser(description="Score dummy, one-feature and rule baselines on a holdout.",
epilog=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
ap.add_argument("csv")
ap.add_argument("--target", required=True)
ap.add_argument("--time-col", help="sort by this column and hold out the latest rows")
ap.add_argument("--test-frac", type=float, default=0.25)
ap.add_argument("--drop", nargs="*", default=[], help="columns that are not features (IDs, leaks)")
ap.add_argument("--rule", help="current business rule for binary targets, e.g. \"calls >= 3 or plan == basic\"")
ap.add_argument("--seed", type=int, default=0)
ap.add_argument("--json", action="store_true")
a = ap.parse_args()
try:
fh = sys.stdin if a.csv == "-" else open(a.csv, newline="", encoding="utf-8-sig")
reader = csv.DictReader(fh)
header = reader.fieldnames or []
rows = [r for r in reader if not missing(r.get(a.target))]
except OSError as e:
print(f"error: {e}", file=sys.stderr)
return 2
if a.target not in header or len(rows) < 20:
print(f"error: target {a.target!r} missing or fewer than 20 labelled rows", file=sys.stderr)
return 2
if a.time_col:
rows.sort(key=lambda r: r.get(a.time_col, ""))
split = "time-ordered (latest rows are the test set)"
else:
random.Random(a.seed).shuffle(rows)
split = f"random (seed {a.seed}); use --time-col if the model will predict forward in time"
cut = int(len(rows) * (1 - a.test_frac))
train, test = rows[:cut], rows[cut:]
classes = Counter(r[a.target].strip() for r in rows)
binary = len(classes) == 2
feats = [c for c in header if c not in (a.target, a.time_col, *a.drop)]
results = []
if binary:
pos = min(classes, key=classes.get)
y_tr = [1 if r[a.target].strip() == pos else 0 for r in train]
y_te = [1 if r[a.target].strip() == pos else 0 for r in test]
prior = sum(y_tr) / len(y_tr)
maj = 1 if prior >= 0.5 else 0
results.append({"baseline": "majority", "detail": f"always {pos if maj else 'not ' + pos}",
**cls_metrics([float(maj)] * len(test), y_te, prior)})
results.append({"baseline": "prior", "detail": f"p = {prior:.3f} for every row",
**cls_metrics([prior] * len(test), y_te, prior)})
m = pick_one_feature(train, feats, y_tr, True)
if m:
results.append({"baseline": "one_r", "detail": m["desc"],
**cls_metrics([m["score"](r) for r in test], y_te, prior)})
if a.rule:
try:
pred = compile_rule(a.rule, header)
except ValueError as e:
print(f"error: {e}", file=sys.stderr)
return 2
results.append({"baseline": "rule", "detail": a.rule,
**cls_metrics([1.0 if pred(r) else 0.0 for r in test], y_te, prior)})
keys = ["accuracy", "balanced_acc", "roc_auc", "pr_auc", "log_loss", "brier"]
head = (f"positive class {pos!r}: train rate {prior:.3f}, test rate {sum(y_te) / len(y_te):.3f}; accuracy at "
f"0.5, balanced_acc at the train rate")
else:
try:
y_tr = [float(r[a.target]) for r in train]
y_te = [float(r[a.target]) for r in test]
except ValueError:
print("error: target is neither binary nor numeric; for multiclass use the majority class "
"per one-vs-rest target", file=sys.stderr)
return 2
mean = sum(y_tr) / len(y_tr)
med = sorted(y_tr)[len(y_tr) // 2]
results.append({"baseline": "mean", "detail": f"{mean:.4g}", **reg_metrics([mean] * len(test), y_te)})
results.append({"baseline": "median", "detail": f"{med:.4g}", **reg_metrics([med] * len(test), y_te)})
m = pick_one_feature(train, feats, y_tr, False)
if m:
results.append({"baseline": "one_r", "detail": m["desc"],
**reg_metrics([m["score"](r) for r in test], y_te)})
keys = ["mae", "rmse", "r2", "wape"]
head = f"numeric target: train mean {mean:.4g}, median {med:.4g}"
out = {"split": split, "train_rows": len(train), "test_rows": len(test), "summary": head, "baselines": results}
if a.json:
print(json.dumps(out, indent=2, default=lambda x: None))
return 0
print(f"Split: {split}; train {len(train)}, test {len(test)}")
print(head)
print(f"\n{'baseline':<10} " + " ".join(f"{k:>12}" for k in keys) + " detail")
for r in results:
print(f"{r['baseline']:<10} " + " ".join(f"{r[k]:>12.4f}" for k in keys) + f" {r['detail']}")
print("\nA model is worth shipping only if it beats the best row here on the primary metric, by more than the "
"noise (compare with a CI or repeated splits).")
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.