2nd/10_Wiki/Topics/AI_and_ML/Imbalanced-Data-Handling.md

id, title, category, status, canonical_id, aliases, duplicate_of, source_trust_level, confidence_score, verification_status, tags, raw_sources, last_reinforced, github_commit, tech_stack

id	title	category	status	canonical_id	aliases	duplicate_of	source_trust_level	confidence_score	verification_status	tags	raw_sources	last_reinforced	github_commit	tech_stack
wiki-2026-0508-imbalanced-data-handling	Imbalanced Data Handling	10_Wiki/Topics	verified	self	imbalanced data class imbalance SMOTE oversampling undersampling class weight	none	A	0.95	applied	machine-learning imbalanced smote oversampling class-weight fraud		2026-05-10	pending	language framework Python imbalanced-learn / scikit-learn

Imbalanced Data Handling

매 한 줄

"매 class distribution 의 의 의 imbalance — 매 majority dominate". 매 fraud, 매 medical rare disease, 매 anomaly 의 common. 매 method: 매 oversample (SMOTE), undersample, class weight, focal loss, threshold tune. 매 evaluation: 매 accuracy 의 useless — PR-AUC, F1, MCC.

매 핵심

매 method

• Resampling:
  • Random oversample (minority).
  • SMOTE: 매 synthetic minority.
  • ADASYN: adaptive.
  • Random undersample (majority).
  • Tomek links: 매 boundary clean.
• Class weight.
• Loss-based: focal loss, weighted CE.
• Threshold tuning: 매 default 0.5 의 X.
• Anomaly detection (1-class).
• Cost-sensitive learning.

매 metric

• PR-AUC (Average Precision).
• F1 / Macro-F1.
• MCC (Matthews Correlation Coefficient).
• Cohen's κ.
• Confusion matrix.

매 응용

1. Fraud detection.
2. Medical (rare disease).
3. Anomaly detection.
4. Customer churn.
5. Click prediction.

💻 패턴

Class weight (sklearn)

from sklearn.utils import class_weight
weights = class_weight.compute_class_weight('balanced', classes=np.unique(y), y=y)
weight_dict = dict(zip(np.unique(y), weights))

from sklearn.linear_model import LogisticRegression
model = LogisticRegression(class_weight=weight_dict).fit(X, y)

SMOTE (imbalanced-learn)

from imblearn.over_sampling import SMOTE
sm = SMOTE(random_state=0)
X_res, y_res = sm.fit_resample(X, y)

SMOTE-NC (mixed numerical + categorical)

from imblearn.over_sampling import SMOTENC
smnc = SMOTENC(categorical_features=[0, 2, 5], random_state=0)
X_res, y_res = smnc.fit_resample(X, y)

Random undersample

from imblearn.under_sampling import RandomUnderSampler
rus = RandomUnderSampler(sampling_strategy=0.5, random_state=0)
X_res, y_res = rus.fit_resample(X, y)

Combined (SMOTE + Tomek)

from imblearn.combine import SMOTETomek
smt = SMOTETomek(random_state=0)
X_res, y_res = smt.fit_resample(X, y)

Pipeline (avoid leakage)

from imblearn.pipeline import Pipeline as ImbPipeline
pipe = ImbPipeline([
    ('scaler', StandardScaler()),
    ('smote', SMOTE()),
    ('clf', LogisticRegression()),
])
# 매 SMOTE applied 매 매 fold (CV-safe)
from sklearn.model_selection import cross_val_score
scores = cross_val_score(pipe, X, y, cv=5, scoring='f1')

Focal loss (PyTorch)

import torch.nn.functional as F
def focal_loss(logits, targets, alpha=0.25, gamma=2.0):
    p = torch.sigmoid(logits)
    p_t = p * targets + (1 - p) * (1 - targets)
    alpha_t = alpha * targets + (1 - alpha) * (1 - targets)
    bce = F.binary_cross_entropy_with_logits(logits, targets, reduction='none')
    return (alpha_t * (1 - p_t) ** gamma * bce).mean()

Threshold tuning

from sklearn.metrics import precision_recall_curve
y_score = model.predict_proba(X_val)[:, 1]
prec, rec, thr = precision_recall_curve(y_val, y_score)
f1_scores = 2 * prec * rec / (prec + rec + 1e-9)
best_thr = thr[f1_scores.argmax()]
y_pred = (y_score > best_thr).astype(int)

XGBoost scale_pos_weight

import xgboost as xgb
ratio = sum(y == 0) / sum(y == 1)
model = xgb.XGBClassifier(scale_pos_weight=ratio).fit(X, y)

Eval metrics (proper)

from sklearn.metrics import classification_report, average_precision_score, matthews_corrcoef, confusion_matrix
print(classification_report(y_val, y_pred))
print(f'PR-AUC: {average_precision_score(y_val, y_score):.3f}')
print(f'MCC: {matthews_corrcoef(y_val, y_pred):.3f}')
print(confusion_matrix(y_val, y_pred))

Cost-sensitive learning

COST_MATRIX = np.array([[0, 1], [10, 0]])  # 매 FN cost = 10x FP

def cost_sensitive_predict(probs, cost):
    expected_cost = probs @ cost
    return expected_cost.argmin(axis=1)

One-class anomaly

from sklearn.ensemble import IsolationForest
iso = IsolationForest(contamination=0.01).fit(X_majority)
anomalies = iso.predict(X_test) == -1

Weighted sampler (PyTorch)

from torch.utils.data import WeightedRandomSampler
class_counts = [sum(y == c) for c in np.unique(y)]
weights = [1.0 / class_counts[label] for label in y]
sampler = WeightedRandomSampler(weights, num_samples=len(y), replacement=True)
loader = DataLoader(dataset, batch_size=32, sampler=sampler)

Stratified split

from sklearn.model_selection import StratifiedKFold
skf = StratifiedKFold(n_splits=5, shuffle=True, random_state=0)
for train_idx, val_idx in skf.split(X, y):
    train_fold(X[train_idx], y[train_idx])

Borderline-SMOTE

from imblearn.over_sampling import BorderlineSMOTE
bsm = BorderlineSMOTE(random_state=0)
X_res, y_res = bsm.fit_resample(X, y)

Calibration check after handling

from sklearn.calibration import calibration_curve
prob_true, prob_pred = calibration_curve(y_val, y_score, n_bins=10)
# 매 oversampling 매 의 의 calibration 의 distort

매 결정 기준

상황	Approach
Mild (< 1:10)	Class weight
Moderate (1:10-1:100)	SMOTE / class weight
Severe (> 1:100)	Anomaly detection / focal
Tabular	XGBoost scale_pos_weight
DL	Focal loss + weighted sampler
Cost varies	Cost-sensitive

기본값: 매 class weight + 매 threshold tune + 매 PR-AUC eval. 매 severe = focal + anomaly detection 의 explore. 매 SMOTE 는 careful (calibration distort).

🔗 Graph

• 부모: Machine-Learning · Data-Preprocessing
• 변형: SMOTE · Focal-Loss
• 응용: Anomaly-Detection

🤖 LLM 활용

언제: 매 fraud / medical / churn / anomaly. 언제 X: 매 balanced.

❌ 안티패턴

• Accuracy metric on imbalanced: 매 misleading.
• SMOTE before train/val split: 매 leakage.
• No threshold tune: 매 default 0.5 의 wrong.
• Aggressive oversample: 매 calibration 의 break.
• Ignore minority cost: 매 FN expensive.

🧪 검증 / 중복

• Verified (Chawla SMOTE 2002, He & Garcia review 2009, Lin focal 2017).
• 신뢰도 A.

🕓 Changelog

날짜	변경
2026-05-08	Phase 1
2026-05-10	Manual cleanup — methods + 매 SMOTE / focal / threshold / scale_pos_weight code