【保存版】ハイパーパラメータ入門：Grid/Random/Optunaの実務チューニング完全ガイド

import pandas as pd
from sklearn.datasets import load_breast_cancer
from sklearn.model_selection import train_test_split, StratifiedKFold
from sklearn.pipeline import Pipeline
from sklearn.preprocessing import StandardScaler
from sklearn.compose import ColumnTransformer
from sklearn.impute import SimpleImputer

Xy = load\_breast\_cancer(as\_frame=True)
X, y = Xy.data, Xy.target
X\_tr, X\_te, y\_tr, y\_te = train\_test\_split(X, y, test\_size=0.2, stratify=y, random\_state=42)

num\_cols = X.columns
prep = ColumnTransformer(\[
("num", Pipeline(\[("imp", SimpleImputer()), ("sc", StandardScaler())]), num\_cols)
])
cv = StratifiedKFold(n\_splits=5, shuffle=True, random\_state=42)

from sklearn.linear_model import LogisticRegression
from sklearn.metrics import roc_auc_score

logit = Pipeline(\[("prep", prep), ("clf", LogisticRegression(max\_iter=1000, n\_jobs=-1, random\_state=42))])
logit.fit(X\_tr, y\_tr)
auc = roc\_auc\_score(y\_te, logit.predict\_proba(X\_te)\[:,1])
print(f"AUC(base-logit)={auc:.3f}")

from sklearn.model_selection import GridSearchCV
from sklearn.ensemble import RandomForestClassifier

rf = Pipeline(\[("prep", prep), ("clf", RandomForestClassifier(random\_state=42, n\_jobs=-1))])
param\_grid = {
"clf\_\_n\_estimators": \[200, 400, 800],
"clf\_\_max\_depth": \[None, 8, 12, 16],
"clf\_\_min\_samples\_split": \[2, 5, 10]
}
search = GridSearchCV(rf, param\_grid, scoring="roc\_auc", cv=cv, n\_jobs=-1)
search.fit(X\_tr, y\_tr)
print("best:", search.best\_params\_, "AUC=", search.best\_score\_)

from sklearn.model_selection import RandomizedSearchCV
from scipy.stats import randint, uniform

param\_dist = {
"clf\_\_n\_estimators": randint(200, 1000),
"clf\_\_max\_depth": \[None] + list(range(4, 21)),
"clf\_\_min\_samples\_split": randint(2, 20),
"clf\_\_min\_samples\_leaf": randint(1, 10),
"clf\_\_max\_features": uniform(0.3, 0.7)  # 0.3〜1.0
}
rand = RandomizedSearchCV(rf, param\_distributions=param\_dist, n\_iter=40,
scoring="roc\_auc", cv=cv, random\_state=42, n\_jobs=-1)
rand.fit(X\_tr, y\_tr)
print("best:", rand.best\_params\_, "AUC=", rand.best\_score\_)

from sklearn.metrics import precision_recall_curve, confusion_matrix
import numpy as np

best = rand.best\_estimator\_.fit(X\_tr, y\_tr)
proba = best.predict\_proba(X\_te)\[:,1]

# 例：上位5%のみ営業（架電枠）

k = int(len(proba) \* 0.05)
idx = np.argsort(-proba)\[:k]
recall\_at\_k = y\_te.iloc\[idx].mean()  # Positive率
print("Recall\@5%:", float(recall\_at\_k))

import json, time
rec = {"when": time.strftime("%Y-%m-%d %H:%M"),
       "best_params": rand.best_params_,
       "cv_auc": float(rand.best_score_)}
with open("search_record.json", "w", encoding="utf-8") as f:
    json.dump(rec, f, ensure_ascii=False, indent=2)

# pip install optuna
import optuna
from sklearn.metrics import roc_auc_score
from sklearn.ensemble import RandomForestClassifier
from sklearn.model_selection import cross_val_score

rf0 = Pipeline(\[("prep", prep), ("clf", RandomForestClassifier(random\_state=42, n\_jobs=-1))])

def objective(trial: optuna.Trial) -> float:
params = {
"clf\_\_n\_estimators": trial.suggest\_int("clf\_\_n\_estimators", 200, 1000),
"clf\_\_max\_depth": trial.suggest\_int("clf\_\_max\_depth", 4, 20),
"clf\_\_min\_samples\_split": trial.suggest\_int("clf\_\_min\_samples\_split", 2, 20),
"clf\_\_min\_samples\_leaf": trial.suggest\_int("clf\_\_min\_samples\_leaf", 1, 10),
"clf\_\_max\_features": trial.suggest\_float("clf\_\_max\_features", 0.3, 1.0)
}
model = rf0.set\_params(\*\*params)
scores = cross\_val\_score(model, X\_tr, y\_tr, scoring="roc\_auc", cv=cv, n\_jobs=-1)
return scores.mean()

study = optuna.create\_study(direction="maximize")
study.optimize(objective, n\_trials=40)
print("best:", study.best\_params, "AUC=", study.best\_value)

from sklearn.model_selection import TimeSeriesSplit, GridSearchCV
from sklearn.linear_model import Ridge

cv\_ts = TimeSeriesSplit(n\_splits=5)
reg = Pipeline(\[("prep", prep), ("reg", Ridge())])
param\_grid = {"reg\_\_alpha": \[0.1, 0.3, 1.0, 3.0, 10.0]}
search = GridSearchCV(reg, param\_grid, scoring="neg\_mean\_absolute\_error", cv=cv\_ts, n\_jobs=-1)
search.fit(X\_tr, y\_tr)
print("best:", search.best\_params\_, "MAE=", -search.best\_score\_)

from sklearn.metrics import roc_auc_score, confusion_matrix, roc_curve

best = search.best\_estimator\_.fit(X\_tr, y\_tr)
proba = best.predict\_proba(X\_te)\[:,1]
print("TEST AUC:", roc\_auc\_score(y\_te, proba))

# 例：Youdenで閾値決定

fpr, tpr, thr = roc\_curve(y\_te, proba)
thr\_opt = thr\[(tpr - fpr).argmax()]
cm = confusion\_matrix(y\_te, (proba >= thr\_opt).astype(int))
print(cm, "threshold=", float(thr\_opt))

import pandas as pd
cv_results = pd.DataFrame(search.cv_results_)
cols = ["mean_test_score","std_test_score","params"]
cv_results[cols].to_csv("grid_results.csv", index=False)