#!/usr/bin/env python3
"""Learn a small, human-readable rule set from a CSV and say how much signal it captures.
Fits a shallow binary decision tree (CART, Gini, depth <= --max-depth, every leaf has at
least --min-leaf rows) in pure Python, then prints:
- the rules as IF ... THEN lines with each leaf's positive rate, size and lift
- holdout ROC-AUC, PR-AUC and precision/recall when flagging leaves above --flag-rate
- a SQL CASE expression you can deploy as a rule, and the same rule for --rule input
- optional comparison with your ML model's holdout metric (--ml-auc)
If a depth-2/3 tree gets close to the model, rules may be the better product: cheaper,
auditable and easy to change. If the gap is large, the signal is in interactions or many
weak features and ML earns its cost.
Usage:
python rule_learner.py data.csv --target churned --drop customer_id --time-col signup_date
python rule_learner.py data.csv --target is_fraud --max-depth 2 --min-leaf 30 --ml-auc 0.91
python rule_learner.py data.csv --target churned --json
Exit codes: 0 fitted, 2 usage/read error or target not binary.
"""
import argparse
import csv
import json
import math
import random
import sys
from collections import Counter, defaultdict
MISSING = {"", "na", "n/a", "nan", "null", "none", "?"}
def num(v):
try:
x = float(v)
return x if math.isfinite(x) else None
except (TypeError, ValueError):
return None
def gini(pos, n):
if n == 0:
return 0.0
p = pos / n
return 2 * p * (1 - p)
def candidate_splits(rows, ys, col, numeric):
"""Yield (description, predicate) pairs for one column."""
if numeric:
xs = sorted({num(r[col]) for r in rows if num(r[col]) is not None})
if len(xs) > 32:
xs = sorted({xs[int(len(xs) * q / 32)] for q in range(1, 32)})
for x in xs[1:]:
yield (col, ">=", x), (lambda r, c=col, t=x: (num(r[c]) is not None and num(r[c]) >= t))
else:
for level, cnt in Counter(r[col] for r in rows).most_common(20):
if cnt >= 2:
yield (col, "==", level), (lambda r, c=col, lv=level: r[c] == lv)
def best_split(rows, ys, cols, kinds, min_leaf):
n, pos = len(ys), sum(ys)
parent = gini(pos, n)
best = None
for col in cols:
for cond, pred in candidate_splits(rows, ys, col, kinds[col]):
lp = ln = 0
for r, y in zip(rows, ys):
if pred(r):
ln += 1
lp += y
rn = n - ln
if ln < min_leaf or rn < min_leaf:
continue
gain = parent - (ln / n) * gini(lp, ln) - (rn / n) * gini(pos - lp, rn)
if best is None or gain > best[0]:
best = (gain, cond, pred)
return best
def grow(rows, ys, cols, kinds, depth, max_depth, min_leaf):
node = {"n": len(ys), "pos": sum(ys)}
if depth < max_depth and 0 < node["pos"] < node["n"]:
s = best_split(rows, ys, cols, kinds, min_leaf)
if s and s[0] > 1e-4:
_, cond, pred = s
left = [(r, y) for r, y in zip(rows, ys) if pred(r)]
right = [(r, y) for r, y in zip(rows, ys) if not pred(r)]
node.update(cond=cond, pred=pred,
yes=grow([r for r, _ in left], [y for _, y in left], cols, kinds, depth + 1, max_depth, min_leaf),
no=grow([r for r, _ in right], [y for _, y in right], cols, kinds, depth + 1, max_depth, min_leaf))
return node
def leaves(node, path=()):
if "cond" not in node:
yield path, node
return
yield from leaves(node["yes"], path + ((node["cond"], True),))
yield from leaves(node["no"], path + ((node["cond"], False),))
def score(node, r, prior):
while "cond" in node:
node = node["yes"] if node["pred"](r) else node["no"]
return (node["pos"] + 2 * prior) / (node["n"] + 2)
def simplify(path):
"""Merge a root-to-leaf path into one interval per numeric column and the categorical tests.
Missing numeric values fail every '>=' test, so they follow the 'no' branches: a column
with no lower bound on the path also admits NULL.
"""
lo, hi, cats = {}, {}, []
for (col, op, val), truth in path:
if op == ">=":
if truth:
lo[col] = max(lo.get(col, val), val)
else:
hi[col] = min(hi.get(col, val), val)
else:
cats.append((col, val, truth))
parts = []
for col in dict.fromkeys(list(lo) + list(hi)):
parts.append(("num", col, lo.get(col), hi.get(col)))
return parts + [("cat", col, val, truth) for col, val, truth in cats]
def rule_text(parts, sql=False):
out = []
for kind, col, a, b in parts:
if kind == "num":
lo, hi = a, b
if sql:
if lo is not None:
out.append(f"{col} >= {lo:g}" + (f" AND {col} < {hi:g}" if hi is not None else ""))
else:
out.append(f"({col} < {hi:g} OR {col} IS NULL)")
else:
out.append(f"{lo:g} <= {col} < {hi:g}" if lo is not None and hi is not None else
f"{col} >= {lo:g}" if lo is not None else f"{col} < {hi:g} (or missing)")
else:
val, truth = a, b
if sql:
q = "'" + str(val).replace("'", "''") + "'"
out.append(f"{col} = {q}" if truth else f"({col} <> {q} OR {col} IS NULL)")
else:
out.append(f"{col} == {val!r}" if truth else f"{col} != {val!r}")
return " AND ".join(out)
def 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):
g = defaultdict(lambda: [0, 0])
for v, t in zip(s, y):
g[v][0] += t
g[v][1] += 1
P, tp, seen, tot = sum(y), 0, 0, 0.0
for v in sorted(g, reverse=True):
tp += g[v][0]
seen += g[v][1]
tot += g[v][0] / P * tp / seen
return tot
def main():
ap = argparse.ArgumentParser(description="Learn a shallow readable rule set and measure it on a holdout.",
epilog=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
ap.add_argument("csv")
ap.add_argument("--target", required=True)
ap.add_argument("--drop", nargs="*", default=[])
ap.add_argument("--time-col", help="hold out the latest rows by this column (else seeded random)")
ap.add_argument("--max-depth", type=int, default=3)
ap.add_argument("--min-leaf", type=int, default=20)
ap.add_argument("--test-frac", type=float, default=0.3)
ap.add_argument("--flag-rate", type=float, help="flag leaves whose positive rate is >= this (default 2x prior)")
ap.add_argument("--ml-auc", type=float, help="your ML model's ROC-AUC on the same holdout, for the comparison")
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 (r.get(a.target) or "").strip().lower() not in MISSING]
except OSError as e:
print(f"error: {e}", file=sys.stderr)
return 2
classes = Counter(r[a.target].strip() for r in rows)
if a.target not in header or len(classes) != 2:
print(f"error: target {a.target!r} must exist and have exactly 2 values (found {len(classes)})", file=sys.stderr)
return 2
pos_label = min(classes, key=classes.get)
if a.time_col:
rows.sort(key=lambda r: r.get(a.time_col, ""))
else:
random.Random(a.seed).shuffle(rows)
cut = int(len(rows) * (1 - a.test_frac))
train, test = rows[:cut], rows[cut:]
y_tr = [1 if r[a.target].strip() == pos_label else 0 for r in train]
y_te = [1 if r[a.target].strip() == pos_label else 0 for r in test]
cols = [c for c in header if c not in (a.target, a.time_col, *a.drop)]
kinds = {}
for c in cols:
vals = [r[c] for r in train if (r[c] or "").strip().lower() not in MISSING]
kinds[c] = bool(vals) and sum(num(v) is not None for v in vals) / len(vals) > 0.95
cols = [c for c in cols if kinds[c] or len({r[c] for r in train}) <= 50] # skip ID-like text
tree = grow(train, y_tr, cols, kinds, 0, a.max_depth, a.min_leaf)
prior = sum(y_tr) / len(y_tr)
flag = a.flag_rate if a.flag_rate is not None else min(0.99, 2 * prior)
s_te = [score(tree, r, prior) for r in test]
flagged = [s >= flag for s in s_te]
tp = sum(1 for f, y in zip(flagged, y_te) if f and y)
nf = sum(flagged)
res = {"positive_label": pos_label, "train_rows": len(train), "test_rows": len(test), "train_rate": prior,
"test_auc": auc(s_te, y_te), "test_pr_auc": avg_precision(s_te, y_te), "flag_rate": flag,
"flagged": nf, "precision": tp / nf if nf else 0.0, "recall": tp / sum(y_te) if sum(y_te) else 0.0,
"rules": []}
sql_when = []
for path, leaf in sorted(leaves(tree), key=lambda pl: -(pl[1]["pos"] / pl[1]["n"])):
rate = leaf["pos"] / leaf["n"]
parts = simplify(path)
text = rule_text(parts) or "(all rows)"
res["rules"].append({"if": text, "rows": leaf["n"], "positive_rate": rate, "lift": rate / prior if prior else 0,
"flag": rate >= flag})
if rate >= flag:
sql_when.append("(" + (rule_text(parts, sql=True) or "1 = 1") + ")")
res["sql"] = ("CASE WHEN " + "\n OR ".join(sql_when) + " THEN 1 ELSE 0 END AS rule_flag") if sql_when else None
if a.ml_auc:
gap = a.ml_auc - res["test_auc"]
res["ml_auc"], res["gap"] = a.ml_auc, gap
res["verdict"] = ("rules are competitive (gap < 0.03 AUC): prefer rules unless ML adds other value" if gap < 0.03
else "ML adds a moderate lift (0.03-0.08 AUC): hybrid or ML if the lift pays its cost"
if gap < 0.08 else "ML clearly ahead (> 0.08 AUC): the signal is beyond a few rules")
if a.json:
out = dict(res)
print(json.dumps(out, indent=2))
return 0
print(f"Rule tree (depth <= {a.max_depth}, min leaf {a.min_leaf}) for {a.target} == {pos_label!r}; "
f"train rate {prior:.3f}")
for rule in res["rules"]:
mark = "FLAG" if rule["flag"] else " "
print(f" {mark} IF {rule['if']}\n THEN rate {rule['positive_rate']:.3f} (n={rule['rows']}, "
f"lift {rule['lift']:.2f}x)")
print(f"\nHoldout ({len(test)} rows): ROC-AUC {res['test_auc']:.3f} PR-AUC {res['test_pr_auc']:.3f} "
f"flagging leaves >= {flag:.3f}: {nf} flagged, precision {res['precision']:.3f}, recall {res['recall']:.3f}")
if a.ml_auc:
print(f"ML model AUC {a.ml_auc:.3f} vs rules {res['test_auc']:.3f} (gap {res['gap']:+.3f}): {res['verdict']}")
if res["sql"]:
print("\nDeployable rule:\n" + res["sql"])
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.