#!/usr/bin/env python3
"""Compute the full metric panel from a predictions CSV, at the operating point you care about.
Binary (columns y_true in {0,1} and y_score in [0,1]):
ranking ROC-AUC, PR-AUC (average precision), lift and precision/recall at top-K
probability log loss, Brier, expected calibration error (ECE, 10 equal-width bins)
threshold confusion matrix, precision, recall, F1, F-beta, balanced accuracy at --threshold
cost with --cost-fp/--cost-fn: expected cost per row at --threshold and the
cost-minimising threshold found by scanning every distinct score
Regression (columns y_true and y_pred):
MAE, RMSE, MedAE, bias (mean error), R2, WAPE, MAPE (skipping zeros, with a count), sMAPE,
and per-segment MAE / bias with --segment
Usage:
python metric_report.py preds.csv # binary, columns y_true,y_score
python metric_report.py preds.csv --k 50 --threshold 0.3 --beta 2
python metric_report.py preds.csv --cost-fp 8 --cost-fn 600
python metric_report.py reg.csv --regression --pred-col y_pred --segment segment
python metric_report.py preds.csv --json
Exit codes: 0 computed, 2 usage or read error (missing columns, non-numeric values, one class only).
"""
import argparse
import csv
import json
import math
import sys
from collections import defaultdict
def read(path, cols):
fh = sys.stdin if path == "-" else open(path, newline="", encoding="utf-8-sig")
reader = csv.DictReader(fh)
missing = [c for c in cols if c and c not in (reader.fieldnames or [])]
if missing:
raise ValueError(f"missing column(s) {missing}; found {reader.fieldnames}")
return list(reader)
def auc(s, y):
pos = sum(y)
neg = len(y) - pos
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):
groups = defaultdict(lambda: [0, 0])
for v, t in zip(s, y):
groups[v][0] += t
groups[v][1] += 1
pos, tp, seen, total = sum(y), 0, 0, 0.0
for v in sorted(groups, reverse=True):
h, n = groups[v]
tp += h
seen += n
total += h / pos * tp / seen
return total
def confusion(s, y, thr):
tp = sum(1 for v, t in zip(s, y) if v >= thr and t)
fp = sum(1 for v, t in zip(s, y) if v >= thr and not t)
fn = sum(1 for v, t in zip(s, y) if v < thr and t)
return tp, fp, fn, len(y) - tp - fp - fn
def binary_report(rows, a):
try:
y = [int(float(r[a.true_col])) for r in rows]
s = [float(r[a.score_col]) for r in rows]
except ValueError as e:
raise ValueError(f"non-numeric value: {e}")
if set(y) - {0, 1}:
raise ValueError("y_true must be 0/1 for binary metrics (use --regression for numeric targets)")
n, pos = len(y), sum(y)
if pos in (0, n):
raise ValueError("only one class present; ranking metrics are undefined")
prev = pos / n
eps = 1e-15
out = {"rows": n, "positives": pos, "prevalence": prev, "roc_auc": auc(s, y), "pr_auc": avg_precision(s, y)}
if all(0 <= v <= 1 for v in s):
out["log_loss"] = -sum(t * math.log(max(eps, v)) + (1 - t) * math.log(max(eps, 1 - v)) for v, t in zip(s, y)) / n
out["brier"] = sum((v - t) ** 2 for v, t in zip(s, y)) / n
bins = defaultdict(lambda: [0.0, 0.0, 0])
for v, t in zip(s, y):
b = bins[min(9, int(v * 10))]
b[0] += v
b[1] += t
b[2] += 1
out["ece"] = sum(abs(sp - st) for sp, st, c in bins.values()) / n
out["calibration"] = [{"bin": f"{k / 10:.1f}-{(k + 1) / 10:.1f}", "n": c, "mean_pred": sp / c, "frac_pos": st / c}
for k, (sp, st, c) in sorted(bins.items())]
order = sorted(range(n), key=lambda i: -s[i])
k = a.k or max(1, round(prev * n))
hits = sum(y[i] for i in order[:k])
out["at_k"] = {"k": k, "precision": hits / k, "recall": hits / pos, "lift": (hits / k) / prev}
tp, fp, fn, tn = confusion(s, y, a.threshold)
prec = tp / (tp + fp) if tp + fp else 0.0
rec = tp / pos
b2 = a.beta ** 2
out["at_threshold"] = {
"threshold": a.threshold, "tp": tp, "fp": fp, "fn": fn, "tn": tn, "precision": prec, "recall": rec,
"f1": 2 * prec * rec / (prec + rec) if prec + rec else 0.0,
f"f{a.beta:g}": (1 + b2) * prec * rec / (b2 * prec + rec) if prec + rec else 0.0,
"balanced_accuracy": (rec + tn / (tn + fp)) / 2 if tn + fp else rec,
"flagged_share": (tp + fp) / n}
if a.cost_fp is not None and a.cost_fn is not None:
def cost(thr):
tp_, fp_, fn_, _ = confusion(s, y, thr)
return (fp_ * a.cost_fp + fn_ * a.cost_fn) / n
best = min(sorted(set(s)) + [float("inf")], key=cost)
out["cost"] = {"cost_fp": a.cost_fp, "cost_fn": a.cost_fn, "at_threshold": cost(a.threshold),
"best_threshold": best, "best_cost": cost(best), "flag_nobody": pos * a.cost_fn / n,
"flag_everybody": (n - pos) * a.cost_fp / n,
"bayes_threshold_if_calibrated": a.cost_fp / (a.cost_fp + a.cost_fn)}
deciles = []
for d in range(10):
part = order[d * n // 10:(d + 1) * n // 10]
if part:
deciles.append({"decile": d + 1, "n": len(part), "positive_rate": sum(y[i] for i in part) / len(part)})
out["deciles"] = deciles
return out
def regression_report(rows, a):
try:
y = [float(r[a.true_col]) for r in rows]
p = [float(r[a.pred_col]) for r in rows]
except ValueError as e:
raise ValueError(f"non-numeric value: {e}")
n = len(y)
err = [pi - yi for pi, yi in zip(p, y)]
mean = sum(y) / n
sst = sum((v - mean) ** 2 for v in y)
abs_err = sorted(abs(e) for e in err)
nz = [(abs(e) / abs(t)) for e, t in zip(err, y) if t != 0]
out = {"rows": n, "mae": sum(abs_err) / n, "rmse": math.sqrt(sum(e * e for e in err) / n),
"medae": abs_err[n // 2] if n % 2 else (abs_err[n // 2 - 1] + abs_err[n // 2]) / 2,
"bias": sum(err) / n, "r2": 1 - sum(e * e for e in err) / sst if sst else float("nan"),
"wape": sum(abs_err) / sum(abs(v) for v in y) if any(y) else float("nan"),
"mape": sum(nz) / len(nz) if nz else float("nan"), "mape_rows_skipped_zero": n - len(nz),
"smape": sum(2 * abs(e) / (abs(t) + abs(pi)) for e, t, pi in zip(err, y, p) if abs(t) + abs(pi)) / n}
if a.segment:
seg = defaultdict(list)
for r, e, t in zip(rows, err, y):
seg[r.get(a.segment, "")].append((e, t))
out["segments"] = [{"segment": k, "n": len(v), "mae": sum(abs(e) for e, _ in v) / len(v),
"bias": sum(e for e, _ in v) / len(v),
"bias_pct": sum(e for e, _ in v) / sum(t for _, t in v) if sum(t for _, t in v) else None}
for k, v in sorted(seg.items())]
return out
def fmt(v):
return f"{v:.4f}" if isinstance(v, float) else str(v)
def main():
ap = argparse.ArgumentParser(description="Metric panel from a predictions CSV.", epilog=__doc__,
formatter_class=argparse.RawDescriptionHelpFormatter)
ap.add_argument("csv")
ap.add_argument("--regression", action="store_true")
ap.add_argument("--true-col", default="y_true")
ap.add_argument("--score-col", default="y_score")
ap.add_argument("--pred-col", default="y_pred")
ap.add_argument("--threshold", type=float, default=0.5)
ap.add_argument("--beta", type=float, default=2.0, help="F-beta (default 2: recall twice as important)")
ap.add_argument("--k", type=int, help="top-K for precision/recall/lift (default: number of positives)")
ap.add_argument("--cost-fp", type=float)
ap.add_argument("--cost-fn", type=float)
ap.add_argument("--segment", help="regression: column to slice MAE and bias by")
ap.add_argument("--json", action="store_true")
a = ap.parse_args()
try:
cols = [a.true_col, a.pred_col if a.regression else a.score_col, a.segment]
rows = read(a.csv, cols)
if not rows:
raise ValueError("no rows")
out = regression_report(rows, a) if a.regression else binary_report(rows, a)
except (OSError, ValueError) as e:
print(f"error: {e}", file=sys.stderr)
return 2
if a.json:
print(json.dumps(out, indent=2))
return 0
if a.regression:
print(f"Regression metrics on {out['rows']} rows")
for k in ("mae", "rmse", "medae", "bias", "r2", "wape", "mape", "smape"):
print(f" {k:<8} {fmt(out[k])}")
if out["mape_rows_skipped_zero"]:
print(f" (MAPE skipped {out['mape_rows_skipped_zero']} rows with y_true = 0)")
for sgm in out.get("segments", []):
bp = "" if sgm["bias_pct"] is None else f" ({sgm['bias_pct']:+.1%})"
print(f" segment {sgm['segment']:<12} n={sgm['n']:<5} MAE {sgm['mae']:.4g} bias {sgm['bias']:+.4g}{bp}")
return 0
print(f"Binary metrics on {out['rows']} rows, {out['positives']} positives (prevalence {out['prevalence']:.2%})")
print(f" ROC-AUC {out['roc_auc']:.4f} PR-AUC {out['pr_auc']:.4f} (random = {out['prevalence']:.4f})")
if "log_loss" in out:
print(f" log loss {out['log_loss']:.4f} Brier {out['brier']:.4f} ECE {out['ece']:.4f}")
k = out["at_k"]
print(f" top-{k['k']}: precision {k['precision']:.3f} recall {k['recall']:.3f} lift {k['lift']:.2f}x")
t = out["at_threshold"]
fb = [x for x in t if x.startswith("f") and x not in ("f1", "fp", "fn", "flagged_share")][0]
print(f" @ {t['threshold']:g}: TP {t['tp']} FP {t['fp']} FN {t['fn']} TN {t['tn']} precision {t['precision']:.3f} "
f"recall {t['recall']:.3f} F1 {t['f1']:.3f} {fb.upper()} {t[fb]:.3f} bal.acc {t['balanced_accuracy']:.3f} "
f"flagged {t['flagged_share']:.1%}")
if "cost" in out:
c = out["cost"]
print(f" cost/row @ {t['threshold']:g}: {c['at_threshold']:.4f} best threshold {c['best_threshold']:.4g} -> "
f"{c['best_cost']:.4f} (flag nobody {c['flag_nobody']:.4f}, flag everybody {c['flag_everybody']:.4f}; "
f"calibrated Bayes threshold {c['bayes_threshold_if_calibrated']:.4f})")
print(" positive rate by score decile: " + " ".join(f"{d['positive_rate']:.2f}" for d in out["deciles"]))
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.