まとめ
- Balanced Accuracy | Evaluating imbalanced datasetsの概要を押さえ、評価対象と読み取り方を整理します。
- Python 3.13 のコード例で算出・可視化し、手順と実務での確認ポイントを確認します。
- 図表や補助指標を組み合わせ、モデル比較や閾値調整に活かすヒントをまとめます。
1. Definition #
Balanced Accuracy is the mean of the true-positive rate (TPR) and the true-negative rate (TNR): \mathrm{Balanced\ Accuracy} = \frac{1}{2}\left(\frac{TP}{TP + FN} + \frac{TN}{TN + FP}\right) For multiclass problems you average the recall of each class in the same spirit.
2. Implementation in Python 3.13 #
python --version # e.g. Python 3.13.0
pip install scikit-learn matplotlib
We reuse the random-forest classifier from the Accuracy article and print both metrics side by side. The bar chart is saved at static/images/eval/classification/accuracy/accuracy_vs_balanced.png, so generate_eval_assets.py can regenerate it whenever you update the notebook.
import matplotlib.pyplot as plt
import numpy as np
from pathlib import Path
from sklearn.datasets import load_breast_cancer
from sklearn.ensemble import RandomForestClassifier
from sklearn.metrics import accuracy_score, balanced_accuracy_score
from sklearn.model_selection import train_test_split
from sklearn.pipeline import make_pipeline
from sklearn.preprocessing import StandardScaler
X, y = load_breast_cancer(return_X_y=True)
X_train, X_test, y_train, y_test = train_test_split(
X, y, test_size=0.25, random_state=42, stratify=y
)
pipeline = make_pipeline(
StandardScaler(),
RandomForestClassifier(random_state=42, n_estimators=300),
)
pipeline.fit(X_train, y_train)
y_pred = pipeline.predict(X_test)
acc = accuracy_score(y_test, y_pred)
bal_acc = balanced_accuracy_score(y_test, y_pred)
print(f"Accuracy: {acc:.3f}, Balanced Accuracy: {bal_acc:.3f}")
Balanced Accuracy weights each class equally by averaging the recall per class.
3. When to prefer Balanced Accuracy #
- Strong class imbalance – plain Accuracy only reflects the majority class, while Balanced Accuracy keeps minority recall visible.
- Model comparison – when benchmark teams submit models on skewed data, Balanced Accuracy makes their performance differences more honest.
- Threshold tuning – combine it with precision/recall plots to see whether both classes remain detectable at your chosen threshold.
4. Companion metrics #
| Metric | Measures | Caveat on imbalanced data |
|---|---|---|
| Accuracy | Overall hit rate | Dominated by the majority class |
| Recall / Sensitivity | Detection rate per class | Requires separate reporting for each class |
| Balanced Accuracy | Mean recall across classes | Highlights minority-class recall loss |
| Macro F1 | Harmonic mean of precision & recall (per class) | Useful when precision also matters |
Summary #
- Balanced Accuracy is the average of per-class recall, making it well suited to imbalanced datasets.
- In Python 3.13, alanced_accuracy_score gives you the value in one line; compare it with Accuracy to show stakeholders the difference.
- Combine it with precision, recall, and F1 metrics to decide how much weight to give each class when evaluating models.