#!/usr/bin/env python3
"""Analyse hyperparameter search results: what mattered, whether the range was right, and what to pick.
Reads scikit-learn `cv_results_` exported to CSV (param_* columns, mean_test_score, std_test_score)
or an Optuna `study.trials_dataframe()` CSV (params_* columns, value, state). Reports:
best top configurations with mean ± std
boundary numeric parameters whose best values sit in the lowest/highest 15% of the searched
range: the optimum may lie outside - widen the range
sensitivity score range across quartile bins of each numeric parameter (catches peaked effects)
plus Spearman for direction (log scale when values span > 1 decade); mean score per
level for categorical parameters
1-SE set configurations within one standard error of the best (sklearn: std / sqrt(folds));
among them the "simplest" by --simpler rules (e.g. smaller max_leaf_nodes, larger C)
progress best-so-far curve by trial order (Optuna number or row order): flat for the last half
means more trials are unlikely to help
Use --lower-better for losses (Optuna minimise studies are detected from negative improvements
only when you pass it).
Usage:
python analyze_search.py cv_results.csv --folds 5
python analyze_search.py trials.csv --simpler "max_leaf_nodes:low" "min_samples_leaf:high"
python analyze_search.py trials.csv --lower-better --json
Exit codes: 0 analysed, 2 usage or read error.
"""
import argparse
import csv
import json
import math
import sys
from collections import defaultdict
def num(v):
try:
x = float(v)
return x if math.isfinite(x) else None
except (TypeError, ValueError):
return None
def spearman(a, b):
def rk(v):
o = sorted(range(len(v)), key=v.__getitem__)
r = [0.0] * len(v)
i = 0
while i < len(o):
j = i
while j + 1 < len(o) and v[o[j + 1]] == v[o[i]]:
j += 1
for k in range(i, j + 1):
r[o[k]] = (i + j) / 2
i = j + 1
return r
ra, rb = rk(a), rk(b)
n = len(a)
ma, mb = sum(ra) / n, sum(rb) / n
va = math.sqrt(sum((x - ma) ** 2 for x in ra))
vb = math.sqrt(sum((x - mb) ** 2 for x in rb))
return 0.0 if not va or not vb else sum((x - ma) * (y - mb) for x, y in zip(ra, rb)) / (va * vb)
def main():
ap = argparse.ArgumentParser(description="Analyse hyperparameter search results.", epilog=__doc__,
formatter_class=argparse.RawDescriptionHelpFormatter)
ap.add_argument("csv")
ap.add_argument("--folds", type=int, default=5, help="CV folds behind std_test_score (for the standard error)")
ap.add_argument("--lower-better", action="store_true")
ap.add_argument("--simpler", nargs="*", default=[], help="param:low|high - which direction is simpler")
ap.add_argument("--top", type=int, default=5)
ap.add_argument("--json", action="store_true")
a = ap.parse_args()
try:
with open(a.csv, newline="", encoding="utf-8-sig") as fh:
rd = csv.DictReader(fh)
cols = rd.fieldnames or []
rows = list(rd)
except OSError as e:
print(f"error: {e}", file=sys.stderr)
return 2
if "mean_test_score" in cols:
fmt, pcols, score_col, std_col = "sklearn", [c for c in cols if c.startswith("param_")], "mean_test_score", "std_test_score"
elif "value" in cols:
fmt, pcols, score_col, std_col = "optuna", [c for c in cols if c.startswith("params_")], "value", None
if "state" in cols:
rows = [r for r in rows if r["state"] == "COMPLETE"]
if "number" in cols:
rows.sort(key=lambda r: int(r["number"]))
else:
print("error: expected scikit-learn cv_results (mean_test_score) or Optuna trials (value)", file=sys.stderr)
return 2
rows = [r for r in rows if num(r[score_col]) is not None]
if len(rows) < 3 or not pcols:
print("error: need >= 3 completed trials with parameter columns", file=sys.stderr)
return 2
sign = -1 if a.lower_better else 1
scores = [num(r[score_col]) for r in rows]
best_i = max(range(len(rows)), key=lambda i: sign * scores[i])
best = scores[best_i]
name = lambda c: c.split("_", 1)[1] if c.startswith(("param_", "params_")) else c
params = {}
for c in pcols:
vals = [r[c] for r in rows]
nums = [num(v) for v in vals]
params[c] = {"numeric": all(x is not None for x in nums) and len(set(nums)) > 2, "values": vals, "nums": nums}
report = {"format": fmt, "trials": len(rows), "best_score": best,
"best_params": {name(c): rows[best_i][c] for c in pcols}, "boundary": [], "sensitivity": [], "levels": {}}
order = sorted(range(len(rows)), key=lambda i: -sign * scores[i])
report["top"] = [{"score": scores[i], "std": num(rows[i][std_col]) if std_col else None,
"params": {name(c): rows[i][c] for c in pcols}} for i in order[:a.top]]
for c, p in params.items():
if p["numeric"]:
xs = p["nums"]
lo, hi = min(xs), max(xs)
logscale = lo > 0 and hi / lo > 10
tx = [math.log10(x) for x in xs] if logscale else xs
tlo, thi = min(tx), max(tx)
bx = tx[best_i]
pos = (bx - tlo) / (thi - tlo) if thi > tlo else 0.5
if pos <= 0.15 or pos >= 0.85:
report["boundary"].append({"param": name(c), "best": xs[best_i], "range": [lo, hi],
"side": "low" if pos <= 0.15 else "high"})
# binned effect: catches non-monotonic (peaked) effects that a rank correlation misses
idx = sorted(range(len(tx)), key=tx.__getitem__)
q = max(1, len(idx) // 4)
bins = [idx[k:k + q] for k in range(0, len(idx), q)]
means = [sum(scores[i] for i in b) / len(b) for b in bins if b]
report["sensitivity"].append({"param": name(c), "spearman": round(spearman(tx, [sign * s for s in scores]), 3),
"effect_range": round(max(means) - min(means), 4),
# peaked only if the best middle bin clearly beats both ends
"peaked": len(means) >= 3 and max(means[1:-1]) - max(means[0], means[-1])
> 0.5 * (max(means) - min(means)),
"scale": "log" if logscale else "linear"})
else:
agg = defaultdict(list)
for v, s in zip(p["values"], scores):
agg[v].append(s)
report["levels"][name(c)] = {k: round(sum(v) / len(v), 4) for k, v in agg.items()}
report["sensitivity"].sort(key=lambda d: -d["effect_range"])
if std_col:
se = (num(rows[best_i][std_col]) or 0) / math.sqrt(a.folds)
within = [i for i in range(len(rows)) if sign * (best - scores[i]) <= se]
rules = [s.split(":") for s in a.simpler if ":" in s]
def simplicity(i):
key = []
for prm, direction in rules:
c = next((pc for pc in pcols if name(pc) == prm), None)
v = num(rows[i][c]) if c else None
key.append((v if direction == "low" else -v) if v is not None else 0)
return key
pick = min(within, key=simplicity) if rules else best_i
report["one_se"] = {"se": se, "count": len(within), "pick": {name(c): rows[pick][c] for c in pcols},
"pick_score": scores[pick]}
running, curve = None, []
for s in scores:
running = s if running is None or sign * s > sign * running else running
curve.append(running)
half = len(curve) // 2
report["progress"] = {"best_at_half": curve[half - 1] if half else curve[-1], "best_at_end": curve[-1],
"improved_in_second_half": sign * (curve[-1] - curve[half - 1]) > 1e-9 if half else True}
if a.json:
print(json.dumps(report, indent=2, default=str))
return 0
print(f"{fmt} search: {len(rows)} trials; best {best:.4f} with " +
", ".join(f"{k}={v}" for k, v in report["best_params"].items()))
print("\ntop configurations:")
for t in report["top"]:
sd = f" ± {t['std']:.4f}" if t["std"] is not None else ""
print(f" {t['score']:.4f}{sd} " + ", ".join(f"{k}={v}" for k, v in t["params"].items()))
print("\nsensitivity (score range across quartile bins of the parameter; Spearman for direction):")
for s in report["sensitivity"]:
shape = "peaked: optimum inside the range" if s["peaked"] else ("higher is better" if s["spearman"] > 0.2 else
"lower is better" if s["spearman"] < -0.2 else "")
print(f" {s['param']:<22} effect {s['effect_range']:.4f} spearman {s['spearman']:+.3f} ({s['scale']} scale)"
+ (f" {shape}" if shape else ""))
for p, lv in report["levels"].items():
print(f" {p:<22} " + ", ".join(f"{k}: {v}" for k, v in lv.items()))
for b in report["boundary"]:
print(f"BOUNDARY: best {b['param']}={b['best']:g} sits at the {b['side']} end of [{b['range'][0]:g}, "
f"{b['range'][1]:g}] - extend the range on that side")
if "one_se" in report:
o = report["one_se"]
print(f"\n{o['count']} configuration(s) within 1 SE ({o['se']:.4f}) of the best; pick: " +
", ".join(f"{k}={v}" for k, v in o["pick"].items()) + f" (score {o['pick_score']:.4f})")
pr = report["progress"]
print(f"best after half the trials {pr['best_at_half']:.4f}, at the end {pr['best_at_end']:.4f}" +
("" if pr["improved_in_second_half"] else " - no gain in the second half: more trials will not help much"))
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.