#!/usr/bin/env python3
"""Recommend a primary metric, secondary metrics and guardrails from the problem's traits.
Answers the question "which number do we optimise and report?" from facts about the
decision rather than habit. Every recommendation comes with the reason and the metric
it replaces when that one would mislead.
Traits (flags):
--task binary | multiclass | multilabel | regression | count | forecast | ranking |
recommendation | anomaly | survival | uplift
--prevalence positive rate (binary/anomaly), e.g. 0.02
--capacity cases acted on per period (top-K decisions); with --population gives K/N
--cost-fp / --cost-fn cost of a false positive / false negative (any unit)
--probabilities the probability itself is used (pricing, expected loss, risk tiers)
--classes-equal multiclass: rare classes matter as much as common ones
--outliers regression: large errors are rare but not to be chased (robust metric)
--big-errors-worse regression: a few large errors are much worse than many small ones
--near-zero regression/forecast: true values can be 0 or close to 0
--many-series forecast: metric must aggregate across series of different scales
--quantiles forecast/regression: intervals or a service level are needed
Usage:
python pick_metric.py --task binary --prevalence 0.02 --capacity 400 --cost-fp 8 --cost-fn 600
python pick_metric.py --task regression --outliers --near-zero
python pick_metric.py --task forecast --many-series --near-zero --quantiles --json
Exit codes: 0 recommendation printed, 2 usage error.
"""
import argparse
import json
import sys
def rec(primary, why, secondary=(), guardrails=(), avoid=(), threshold=None):
return {"primary": primary, "why": why, "secondary": list(secondary), "guardrails": list(guardrails),
"avoid": list(avoid), "threshold": threshold}
def binary(a):
avoid = []
if a.prevalence is not None and a.prevalence < 0.2:
avoid.append(("accuracy", f"at {a.prevalence:.1%} prevalence 'always negative' scores {1 - a.prevalence:.1%}"))
avoid.append(("F1 at 0.5", "0.5 is rarely the operating point; F1 ignores true negatives and costs"))
thr = None
if a.cost_fp is not None and a.cost_fn is not None and a.cost_fp + a.cost_fn > 0:
t = a.cost_fp / (a.cost_fp + a.cost_fn)
thr = (f"with calibrated probabilities act when p > cost_fp / (cost_fp + cost_fn) = {t:.4f}; "
"check calibration first")
if a.capacity:
k = f"top {a.capacity:g}" + (f" of {a.population:g} ({a.capacity / a.population:.1%})" if a.population else "")
primary = f"precision@K and recall@K at K = {k}"
why = "the action handles a fixed number of cases, so only the ranking of the top K matters"
sec = ["PR-AUC (average precision)", "ROC-AUC for overall ranking", "lift@K vs random and vs the rule"]
elif a.cost_fp is not None and a.cost_fn is not None:
primary = "expected cost (or net value) at the cost-optimal threshold"
why = f"errors have known, asymmetric costs (FP {a.cost_fp:g}, FN {a.cost_fn:g})"
sec = ["PR-AUC", "recall and precision at the chosen threshold", "ROC-AUC"]
elif a.prevalence is not None and a.prevalence < 0.1:
primary = "PR-AUC (average precision)"
why = "rare positives: PR-AUC focuses on how well positives are ranked above negatives near the top"
sec = ["recall at a fixed precision (or precision at a fixed recall)", "ROC-AUC"]
avoid.append(("ROC-AUC alone", "looks high on rare-event problems even when precision is poor"))
else:
primary = "ROC-AUC"
why = "balanced-enough classes and no fixed operating point yet: measure ranking quality"
sec = ["PR-AUC", "balanced accuracy at the chosen threshold"]
if a.probabilities:
sec = ["log loss", "Brier score", "calibration curve / ECE"] + sec
why += "; probabilities feed a downstream calculation, so calibration is part of success"
guard = ["metric by segment (fairness, key customer groups)", "score distribution drift (PSI)"]
if a.capacity:
guard.insert(0, "false positive load on the team per period")
return rec(primary, why, sec, guard, avoid, thr)
def multiclass(a):
if a.classes_equal:
return rec("macro-F1", "every class counts equally, so small classes must not be drowned by big ones",
["per-class recall and precision", "confusion matrix of costly swaps", "top-k accuracy"],
["worst-class recall"], [("accuracy", "dominated by frequent classes")])
return rec("accuracy (or weighted-F1)", "class frequency reflects importance",
["macro-F1", "log loss if probabilities route work", "top-k accuracy for human-in-the-loop routing"],
["per-class recall of critical classes"], [("macro-F1 as primary", "over-weights rare classes")])
def regression(a, count=False):
avoid, sec = [], []
if count:
return rec("Poisson deviance (or MAE)", "count target: errors scale with the mean, Poisson deviance "
"matches a Poisson/Tweedie objective",
["MAE", "calibration by predicted decile (mean predicted vs mean actual)"],
["total predicted vs actual (bias)"], [("RMSE alone", "dominated by the largest counts")])
if a.big_errors_worse and not a.outliers:
primary, why = "RMSE", "large errors are disproportionately costly; RMSE squares them"
sec.append("MAE (typical error)")
else:
primary, why = "MAE", "robust, in the target's units, easy to explain"
sec.append("RMSE (sensitivity to large errors)")
if a.outliers:
why += "; outliers should not dominate the metric"
if a.near_zero:
avoid.append(("MAPE", "explodes or is undefined when actuals are near 0"))
sec.append("WAPE = sum|error| / sum|actual| for business reporting")
else:
sec.append("MAPE or WAPE for business reporting")
sec += ["R2 for comparison across models only", "mean error (bias)"]
if a.quantiles:
sec.insert(0, "pinball loss at each quantile; interval coverage")
return rec(primary, why, sec, ["error by segment", "bias over time"], avoid)
def forecast(a):
avoid = []
if a.many_series:
primary = "MASE (or RMSSE) vs seasonal naive, averaged across series"
why = "scale-free across series and anchored to the naive forecast (MASE < 1 beats it)"
else:
primary = "MAE / WAPE on the forecast horizon"
why = "single series: errors in units are interpretable"
sec = ["WAPE weighted by volume or margin", "bias (sum of forecast - actual) / sum of actual"]
if a.near_zero:
avoid.append(("MAPE / sMAPE", "unstable with zeros and intermittent demand"))
if a.quantiles:
sec.insert(0, "pinball loss / CRPS for quantile forecasts; coverage of prediction intervals")
return rec(primary, why, sec, ["stock-out rate or service level", "error by horizon"], avoid)
SIMPLE = {
"multilabel": rec("micro-F1 (or per-label PR-AUC averaged)", "many labels per item; micro-F1 weights by "
"label frequency", ["macro-F1", "Hamming loss", "per-label recall"], ["label cardinality drift"]),
"ranking": rec("NDCG@k (k = what users see)", "graded relevance and position both matter",
["MRR for single-answer tasks", "MAP / recall@k for binary relevance"],
["zero-result rate", "online CTR / conversion"], [("accuracy/AUC per row", "ignores the query grouping")]),
"recommendation": rec("recall@k (or NDCG@k) on a time-based holdout", "users see a short list; "
"measure whether future interactions appear in it",
["coverage of the catalogue", "novelty / popularity bias", "MAP@k"],
["online CTR, conversion, long-term retention"], [("RMSE on ratings", "does not measure top-k quality")]),
"anomaly": rec("precision@k on a reviewed sample (alerts the team can handle)", "labels are scarce; alert "
"budget defines success", ["recall on known incidents", "time to detect", "PR-AUC when labels exist"],
["alert volume per day"], [("accuracy", "anomalies are rare by definition")]),
"survival": rec("concordance index (Harrell's C or Uno's C for heavy censoring)", "ranks who fails first "
"and uses censored rows", ["time-dependent AUC at the decision horizon", "integrated Brier score"],
["calibration at key horizons"], [("accuracy on 'event yes/no'", "ignores censoring")]),
"uplift": rec("Qini coefficient / AUUC on a randomised holdout", "the decision is who to treat; success is "
"incremental outcome, not response", ["uplift at top-K", "incremental conversions per treatment cost"],
["treatment cost", "negative-effect segments"], [("AUC of response", "rewards targeting sure things")]),
}
def main():
ap = argparse.ArgumentParser(description="Recommend ML success metrics from problem traits.", epilog=__doc__,
formatter_class=argparse.RawDescriptionHelpFormatter)
ap.add_argument("--task", required=True, choices=["binary", "multiclass", "multilabel", "regression", "count",
"forecast", "ranking", "recommendation", "anomaly", "survival",
"uplift"])
ap.add_argument("--prevalence", type=float)
ap.add_argument("--capacity", type=float)
ap.add_argument("--population", type=float)
ap.add_argument("--cost-fp", type=float)
ap.add_argument("--cost-fn", type=float)
ap.add_argument("--probabilities", action="store_true")
ap.add_argument("--classes-equal", action="store_true")
ap.add_argument("--outliers", action="store_true")
ap.add_argument("--big-errors-worse", action="store_true")
ap.add_argument("--near-zero", action="store_true")
ap.add_argument("--many-series", action="store_true")
ap.add_argument("--quantiles", action="store_true")
ap.add_argument("--json", action="store_true")
a = ap.parse_args()
if a.prevalence is not None and not 0 < a.prevalence < 1:
print("error: --prevalence must be between 0 and 1", file=sys.stderr)
return 2
r = {"binary": binary, "multiclass": multiclass, "regression": regression, "forecast": forecast,
"count": lambda x: regression(x, count=True)}.get(a.task, lambda x: SIMPLE[a.task])(a)
if a.json:
print(json.dumps({"task": a.task, **r, "avoid": [{"metric": m, "why": w} for m, w in r["avoid"]]}, indent=2))
return 0
print(f"Task: {a.task}")
print(f"PRIMARY {r['primary']}\n why: {r['why']}")
if r["threshold"]:
print(f"THRESHOLD {r['threshold']}")
for s in r["secondary"]:
print(f"secondary {s}")
for g in r["guardrails"]:
print(f"guardrail {g}")
for m, w in r["avoid"]:
print(f"avoid {m}: {w}")
print("\nReport the primary metric with a confidence interval, against the baseline, on a deployment-like split.")
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.