#!/usr/bin/env python3
"""Read a validation curve (train and validation score across one hyperparameter) for under- and overfitting.
Input: CSV with one row per (parameter value, fold), as written by sklearn.model_selection.validation_curve:
max_depth,fold,train_score,val_score
The parameter column is the first column that is not fold/train/val (or --param-col). Rows per value without a
fold column also work. Scores are higher-is-better (scikit-learn convention) unless --lower-better.
Reports for each value: mean train and validation score, validation fold std and the gap, plus a zone:
underfit less complex than the best value and worse than best - 1 SE
sweet spot within 1 SE of the best validation score
overfit more complex than the best value, worse than best - 1 SE, with a larger gap than at the best
Then:
best value with the best mean validation score
1-SE pick the least complex value within one standard error of the best (std / sqrt(folds))
boundary best value at the edge of the swept range: extend the sweep
current with --current, which zone the value in use falls in and which way to move it
Complexity direction (does a larger value mean a more flexible model?) is taken from the parameter name
(max_depth, num_leaves, C, n_estimators... increase; min_samples_leaf, alpha, reg_lambda, n_neighbors...
decrease) or, if unknown or ambiguous (gamma means opposite things in SVMs and XGBoost), from the
training-score trend; override with --complexity.
Usage:
python validation_curve_report.py validation_curve.csv
python validation_curve_report.py vc.csv --param-col param_C --current 100
python validation_curve_report.py vc.csv --complexity decreasing --json
Exit codes: 0 reported, 2 usage or read error.
"""
import argparse
import csv
import json
import math
import re
import sys
from collections import OrderedDict
MORE_COMPLEX = re.compile(r"(max_depth|depth|num_leaves|max_leaf_nodes|n_estimators|iterations|num_boost_round|"
r"^c$|_c$|max_features|degree|n_components|hidden|units|layers|max_bins|"
r"learning_rate|colsample|subsample|epochs)", re.I)
LESS_COMPLEX = re.compile(r"(min_samples_leaf|min_samples_split|min_child|min_data|alpha|lambda|reg_|l2|l1|"
r"n_neighbors|k_neighbors|ccp_alpha|min_impurity|dropout|smoothing|l2_leaf_reg|"
r"min_gain|weight_decay)", re.I)
SKIP = re.compile(r"^(fold|split|cv|seed|repeat)$", re.I)
def num(v):
try:
x = float(v)
return x if math.isfinite(x) else None
except (TypeError, ValueError):
return None
def mean(xs):
return sum(xs) / len(xs)
def std(xs):
if len(xs) < 2:
return 0.0
m = mean(xs)
return math.sqrt(sum((x - m) ** 2 for x in xs) / (len(xs) - 1))
def load(path, param_col, train_col, val_col):
f = sys.stdin if path == "-" else open(path, newline="", encoding="utf-8")
with f:
rows = list(csv.DictReader(f))
if not rows:
raise ValueError("no rows")
cols = list(rows[0])
if param_col is None:
param_col = next((c for c in cols if c not in (train_col, val_col) and not SKIP.match(c)), None)
for c in (param_col, train_col, val_col):
if c is None or c not in cols:
raise ValueError(f"column {c!r} not found (have: {', '.join(cols)})")
groups = OrderedDict()
for r in rows:
t, v = num(r[train_col]), num(r[val_col])
if t is None or v is None:
continue
groups.setdefault(r[param_col], ([], []))
groups[r[param_col]][0].append(t)
groups[r[param_col]][1].append(v)
if len(groups) < 3:
raise ValueError("need at least three parameter values")
keys = list(groups)
numeric = all(num(k) is not None for k in keys)
if numeric:
keys.sort(key=num)
pts = [{"value": k, "train": mean(groups[k][0]), "val": mean(groups[k][1]), "val_std": std(groups[k][1]),
"folds": len(groups[k][1])} for k in keys]
return param_col, numeric, pts
def fmt(x):
return f"{x:.4g}" if abs(x) < 1e5 else f"{x:.0f}"
def analyse(param, numeric, pts, lower=False, complexity=None, current=None):
s = -1.0 if lower else 1.0
for p in pts:
p["gap"] = s * (p["train"] - p["val"])
vals = [s * p["val"] for p in pts]
ib = max(range(len(pts)), key=lambda i: vals[i])
best = pts[ib]
se = best["val_std"] / math.sqrt(best["folds"]) if best["folds"] > 1 else 0.0
source = "given"
if complexity is None:
if MORE_COMPLEX.search(param) and not LESS_COMPLEX.search(param):
complexity, source = "increasing", "parameter name"
elif LESS_COMPLEX.search(param):
complexity, source = "decreasing", "parameter name"
else:
tr = [s * p["train"] for p in pts]
complexity = "increasing" if tr[-1] >= tr[0] else "decreasing"
source = "training-score trend"
order = list(range(len(pts))) if complexity == "increasing" else list(range(len(pts)))[::-1]
rank = {i: r for r, i in enumerate(order)} # 0 = least complex
within = [i for i in range(len(pts)) if vals[i] >= vals[ib] - se]
pick = min(within, key=lambda i: rank[i])
for i, p in enumerate(pts):
if i in within:
p["zone"] = "sweet spot"
elif rank[i] < rank[ib]:
p["zone"] = "underfit"
else:
p["zone"] = "overfit" if p["gap"] > best["gap"] else "worse"
res = {"param": param, "numeric": numeric, "complexity": complexity, "complexity_source": source,
"points": pts, "best": best["value"], "best_val": best["val"], "se": se, "pick_1se": pts[pick]["value"],
"pick_1se_val": pts[pick]["val"], "lower_better": lower, "warnings": []}
if best["folds"] < 2:
res["warnings"].append("one score per value: the 1-SE rule needs fold scores; export per-fold results")
if ib in (0, len(pts) - 1) and numeric:
side = "low" if ib == 0 else "high"
res["warnings"].append(f"best {param}={best['value']} is at the {side} end of the sweep: extend the range "
f"{'below' if side == 'low' else 'above'} it")
over = [p for p in pts if p["zone"] == "overfit"]
under = [p for p in pts if p["zone"] == "underfit"]
if not over and not under:
res["warnings"].append("every value is within 1 SE of the best: this parameter barely matters over this range")
if current is not None:
key = next((p for p in pts if p["value"] == current or (numeric and num(current) is not None
and num(p["value"]) == num(current))), None)
if key is None:
res["warnings"].append(f"--current {current} is not one of the swept values")
else:
z = key["zone"]
move = ("more complex" if z == "underfit" else "less complex" if z in ("overfit", "worse") else None)
if move:
larger = (move == "more complex") == (complexity == "increasing")
move += f" setting ({'larger' if larger else 'smaller'} {param}, towards {pts[pick]['value']})"
res["current"] = {"value": current, "zone": z, "move": move}
return res
def report(r):
lines = [f"parameter: {r['param']} complexity grows with {'larger' if r['complexity'] == 'increasing' else 'smaller'}"
f" values ({r['complexity_source']})",
f"{'value':>10} {'train':>9} {'val':>9} {'val_std':>8} {'gap':>9} zone"]
for p in r["points"]:
mark = " <- best" if p["value"] == r["best"] else (" <- 1-SE pick" if p["value"] == r["pick_1se"] else "")
lines.append(f"{p['value']:>10} {fmt(p['train']):>9} {fmt(p['val']):>9} {fmt(p['val_std']):>8} "
f"{fmt(p['gap']):>9} {p['zone']}{mark}")
lines.append("")
lines.append(f"BEST: {r['param']}={r['best']} (validation {fmt(r['best_val'])}, 1 SE = {fmt(r['se'])})")
lines.append(f"1-SE PICK: {r['param']}={r['pick_1se']} (validation {fmt(r['pick_1se_val'])}): the least complex "
"value statistically tied with the best")
c = r.get("current")
if c:
lines.append(f"CURRENT: {r['param']}={c['value']} is in the {c['zone']} zone"
+ (f": move to a {c['move']}" if c["move"] else ": keep it"))
for w in r["warnings"]:
lines.append(f"WARNING: {w}")
return "\n".join(lines)
def main():
ap = argparse.ArgumentParser(description="Read a validation curve.", epilog=__doc__,
formatter_class=argparse.RawDescriptionHelpFormatter)
ap.add_argument("csv", help="validation curve CSV, or - for stdin")
ap.add_argument("--param-col")
ap.add_argument("--train-col", default="train_score")
ap.add_argument("--val-col", default="val_score")
ap.add_argument("--lower-better", action="store_true")
ap.add_argument("--complexity", choices=["increasing", "decreasing"],
help="whether larger parameter values make the model more flexible")
ap.add_argument("--current", help="value currently in use")
ap.add_argument("--json", action="store_true")
a = ap.parse_args()
try:
param, numeric, pts = load(a.csv, a.param_col, a.train_col, a.val_col)
except (OSError, ValueError) as e:
print(f"error: {e}", file=sys.stderr)
return 2
res = analyse(param, numeric, pts, a.lower_better, a.complexity, a.current)
print(json.dumps(res, indent=2) if a.json else report(res))
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.