무작위 오버샘플링과 SMOTE
학습자는 imbalanced-learn의 RandomOverSampler와 SMOTE를 적용해 소수 클래스를 오버샘플링한 다음, 정밀도와 재현율이 향상되는지 평가합니다.
무작위 오버샘플링과 SMOTE은(는) CoddyKit의 무료 Machine Learning Academy 강의입니다. 이것은 4개 중 2번째 강의입니다. 아래에서 전체 강의를 무료로 읽을 수 있으며, 내장 코드 에디터와 24/7 AI 튜터와 함께 브라우저에서 직접 실습할 수 있습니다. 이 강의는 Machine Learning Academy 학습 경로의 일부이며, 진행 상황이 웹과 CoddyKit 앱에 동기화됩니다. Machine Learning Academy 강의에는 총 4개의 강의가 포함되어 있습니다.
이 강의의 일부는 아직 번역되지 않았으며 영어로 표시됩니다.
Why Oversampling Helps Imbalanced Models
Many classifiers learn a biased decision boundary when training data is dominated by one class — they effectively ignore the minority class. Oversampling increases the representation of the minority class in the training data, forcing the model to pay more attention to it. Two popular approaches are random oversampling (duplicating existing minority samples) and SMOTE (synthesising new ones).
Random Oversampling: Duplicate Minority Samples
RandomOverSampler from the imbalanced-learn library randomly duplicates examples from the minority class until the desired class ratio is achieved. It is simple and effective, but the duplicate samples add no new information — the model may overfit to the repeated minority examples if not regularised carefully.
from imblearn.over_sampling import RandomOverSampler
from sklearn.datasets import make_classification
import numpy as np
X, y = make_classification(n_samples=1000, weights=[0.95, 0.05],
n_features=10, random_state=42)
print('Before resampling:', np.bincount(y))
ros = RandomOverSampler(random_state=42)
X_res, y_res = ros.fit_resample(X, y)
print('After resampling: ', np.bincount(y_res))Installing imbalanced-learn
imbalanced-learn is a scikit-learn compatible library for handling class imbalance. Install it with pip install imbalanced-learn. It follows the sklearn API — samplers have fit_resample instead of fit_transform. It integrates with sklearn Pipelines via imblearn.pipeline.Pipeline, which is a drop-in replacement that supports samplers as pipeline steps.
# pip install imbalanced-learn
from imblearn import __version__ as imb_version
from imblearn.over_sampling import RandomOverSampler, SMOTE
from imblearn.under_sampling import RandomUnderSampler
from imblearn.pipeline import Pipeline as ImbPipeline
print('imbalanced-learn version:', imb_version)SMOTE: Synthetic Minority Oversampling
SMOTE (Synthetic Minority Oversampling TEchnique) creates new synthetic minority examples rather than duplicating existing ones. For each minority sample, SMOTE finds its k nearest minority neighbours and creates new points along the line segments connecting them. These synthetic samples occupy the interior of the minority feature space, giving the classifier richer information about the minority class boundary.
SMOTE in Practice
Use imblearn.over_sampling.SMOTE. Key parameters: k_neighbors (default 5) — how many minority neighbours to consider. The sampling_strategy parameter controls the target ratio: 'auto' oversamples all minority classes to match the majority class count.
from imblearn.over_sampling import SMOTE
from sklearn.datasets import make_classification
import numpy as np
X, y = make_classification(n_samples=1000, weights=[0.95, 0.05],
n_features=10, random_state=42)
print('Before SMOTE:', np.bincount(y))
smote = SMOTE(k_neighbors=5, random_state=42)
X_res, y_res = smote.fit_resample(X, y)
print('After SMOTE:', np.bincount(y_res))
print('New shape:', X_res.shape)Comparing Random Oversampling vs SMOTE
Random oversampling simply copies existing data and is faster; SMOTE synthesises new data and tends to generalise better. However, SMOTE can create unrealistic samples if the minority class is very sparse or if features have complex interactions. On highly non-linear problems, borderline-SMOTE or ADASYN (Adaptive Synthetic Sampling) may work better than standard SMOTE.
from imblearn.over_sampling import RandomOverSampler, SMOTE
from sklearn.linear_model import LogisticRegression
from sklearn.metrics import f1_score
from sklearn.datasets import make_classification
from sklearn.model_selection import train_test_split
from sklearn.preprocessing import StandardScaler
import numpy as np
X, y = make_classification(n_samples=1000, weights=[0.95, 0.05], random_state=42)
X_train, X_test, y_train, y_test = train_test_split(X, y, stratify=y, random_state=42)
sc = StandardScaler()
X_train_s = sc.fit_transform(X_train)
X_test_s = sc.transform(X_test)
for name, sampler in [('None', None), ('ROS', RandomOverSampler(random_state=0)),
('SMOTE', SMOTE(random_state=0))]:
if sampler:
X_r, y_r = sampler.fit_resample(X_train_s, y_train)
else:
X_r, y_r = X_train_s, y_train
lr = LogisticRegression().fit(X_r, y_r)
f1 = f1_score(y_test, lr.predict(X_test_s))
print(f'{name:6s}: F1={f1:.4f}')Using imbalanced-learn Pipeline
The critical rule: apply SMOTE only to the training fold, never to the test fold. The imblearn.pipeline.Pipeline enforces this automatically — SMOTE's fit_resample is called during fit but not during predict or transform. This is why you must use imblearn.pipeline.Pipeline, not sklearn's, when including a sampler.
from imblearn.pipeline import Pipeline as ImbPipeline
from imblearn.over_sampling import SMOTE
from sklearn.preprocessing import StandardScaler
from sklearn.linear_model import LogisticRegression
from sklearn.model_selection import cross_val_score
from sklearn.datasets import make_classification
import numpy as np
X, y = make_classification(n_samples=1000, weights=[0.95, 0.05], random_state=42)
pipe = ImbPipeline([
('sc', StandardScaler()),
('smote', SMOTE(random_state=42)), # applied only during fit
('lr', LogisticRegression())
])
scores = cross_val_score(pipe, X, y, cv=5, scoring='f1')
print(f'CV F1: {np.mean(scores):.4f} +/- {np.std(scores):.4f}')Visualising SMOTE Samples
On a 2D dataset you can visualise where SMOTE places synthetic samples. The new points lie along line segments between existing minority samples in feature space. This confirms that SMOTE is interpolating within the minority class region — not extrapolating outside it.
import matplotlib.pyplot as plt
from imblearn.over_sampling import SMOTE
from sklearn.datasets import make_classification
X, y = make_classification(n_samples=200, weights=[0.9, 0.1],
n_features=2, n_redundant=0,
n_informative=2, random_state=42)
smote = SMOTE(random_state=0)
X_res, y_res = smote.fit_resample(X, y)
fig, (ax1, ax2) = plt.subplots(1, 2, figsize=(10, 4))
for ax, data, labels, title in [(ax1, X, y, 'Before SMOTE'),
(ax2, X_res, y_res, 'After SMOTE')]:
ax.scatter(data[labels==0, 0], data[labels==0, 1], label='Class 0', alpha=0.5)
ax.scatter(data[labels==1, 0], data[labels==1, 1], label='Class 1', alpha=0.5)
ax.set_title(title)
ax.legend()
plt.tight_layout()
plt.show()SMOTE Variants: Borderline and ADASYN
Borderline-SMOTE generates synthetic samples only near the decision boundary (borderline minority samples closest to the majority class), focusing the model's attention where classification is hardest. ADASYN adaptively generates more samples in regions where the class is densest, adding proportionally more synthetic examples where the classifier struggles most.
from imblearn.over_sampling import BorderlineSMOTE, ADASYN
from sklearn.datasets import make_classification
import numpy as np
X, y = make_classification(n_samples=1000, weights=[0.95, 0.05], random_state=42)
print('Before:', np.bincount(y))
bs = BorderlineSMOTE(random_state=0)
X_bs, y_bs = bs.fit_resample(X, y)
print('After BorderlineSMOTE:', np.bincount(y_bs))
adas = ADASYN(random_state=0)
X_ad, y_ad = adas.fit_resample(X, y)
print('After ADASYN:', np.bincount(y_ad))Combining SMOTE with Undersampling (SMOTEENN)
SMOTE alone can introduce noisy samples near the majority-class boundary. SMOTEENN combines SMOTE with Edited Nearest Neighbours (ENN) cleaning: after SMOTE upsamples, ENN removes misclassified samples from both classes near the boundary. The result is a cleaner dataset with both oversampled minority and cleaned majority regions.
from imblearn.combine import SMOTEENN
from sklearn.datasets import make_classification
import numpy as np
X, y = make_classification(n_samples=1000, weights=[0.95, 0.05], random_state=42)
print('Before:', np.bincount(y))
smoteenn = SMOTEENN(random_state=42)
X_res, y_res = smoteenn.fit_resample(X, y)
print('After SMOTEENN:', np.bincount(y_res))Evaluating SMOTE: Precision-Recall Trade-Off
After applying oversampling, always evaluate with metrics sensitive to the minority class — F1-score, ROC-AUC, or PR-AUC. SMOTE typically improves recall (fewer missed positives) but may reduce precision (more false alarms). Use the precision-recall curve to find the threshold that gives your desired operating point for the specific business cost of false positives vs false negatives.
Quick Check
Test your understanding of oversampling and SMOTE from this lesson.
Lesson Recap
In this lesson you learned: RandomOverSampler duplicates minority examples — simple but risks overfitting to repeated data, SMOTE synthesises new minority samples by interpolating between existing ones, giving the classifier richer information, and imblearn.pipeline.Pipeline ensures SMOTE is applied only to training folds, preventing data leakage. Next up we explore undersampling techniques that reduce the majority class instead of growing the minority.
자주 묻는 질문
“무작위 오버샘플링과 SMOTE” 강의는 무료인가요?
네 — “무작위 오버샘플링과 SMOTE” 전체 내용을 이 웹사이트에서 무료로 읽을 수 있습니다. 인터랙티브하게 실습하려면(내장 코드 에디터와 24/7 AI 튜터), CoddyKit PRO로 업그레이드하면 Machine Learning Academy 강의 전체를 잠금 해제할 수 있습니다. Machine Learning Academy 강의에는 총 4개의 강의가 포함되어 있습니다.
“무작위 오버샘플링과 SMOTE”에서 뭘 배우나요?
학습자는 imbalanced-learn의 RandomOverSampler와 SMOTE를 적용해 소수 클래스를 오버샘플링한 다음, 정밀도와 재현율이 향상되는지 평가합니다. 브라우저에서 직접 실행하는 실습 코드로 Machine Learning Academy을(를) 배우며, 24/7 AI 튜터가 강의를 진행하면서 질문에 답변해줍니다.
Machine Learning Academy을(를) 시작하는 데 경험이 필요한가요?
사전 경험은 필요하지 않습니다. CoddyKit의 Machine Learning Academy은(는) 초급자부터 고급 학습자까지를 위해 구성되어 있으므로, 여기서 시작하거나 처음부터 시작할 수 있으며 자신의 속도대로 진행할 수 있습니다. 이것은 4개 중 2번째 강의입니다.
“무작위 오버샘플링과 SMOTE” 강의는 얼마나 걸리나요?
대부분의 CoddyKit 강의는 약 5~10분이 소요됩니다. 각 강의는 간결하고 인터랙티브하여 꾸준한 진행이 가능하며, 웹과 앱에서 중단한 부분부터 바로 시작할 수 있습니다.
이 Machine Learning Academy 강의에서 코드를 작성하고 실행할 수 있나요?
네. 모든 Machine Learning Academy 강의에는 내장 코드 에디터가 포함되어 있으므로, 브라우저에서 바로 실제 코드를 작성하고 실행한 후 즉시 AI 피드백을 받을 수 있습니다 — 로컬 설정이 필요 없습니다.