2nd/10_Wiki/Topics/AI_and_ML/One-Hot-Encoding.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-one-hot-encoding	One-Hot Encoding	10_Wiki/Topics	verified	self	One-Hot OHE Indicator-Encoding Dummy-Encoding	none	A	0.95	applied	feature-engineering categorical preprocessing sklearn pandas		2026-05-10	pending	language framework python sklearn-pandas

One-Hot Encoding

매 한 줄

"매 categorical value → orthogonal binary vector". One-hot encoding 은 K 개 category 를 K 개 0/1 column 으로 펼치는 매 가장 단순한 categorical → numeric 변환. 매 linear model / tree-based model 의 default, 그러나 high-cardinality 에서는 target / hash encoding 으로 교체.

매 핵심

매 정의

• category set {A, B, C} → vectors (1,0,0), (0,1,0), (0,0,1).
• ordinal encoding (0,1,2) 와 달리 순서 가정 없음.
• linear / kernel model 의 가정 (numeric distance) 을 깨지 않음.

매 dummy variable trap

• K columns → 1 redundant (sum=1 의 collinearity).
• linear regression 의 unregularized 경우 → drop_first=True.
• tree / regularized model (Lasso, Ridge) → 매 전체 K 유지 가능.

매 cardinality 의 문제

• high-cardinality (>50): sparse matrix 폭발, leak 위험.
• 대안: target / mean encoding, hashing trick, embedding.

매 응용

1. tabular ML 의 categorical preprocessing.
2. NLP token → vocab vector (sparse).
3. RL action / state space 의 discrete encoding.

💻 패턴

sklearn OneHotEncoder

from sklearn.preprocessing import OneHotEncoder
import numpy as np

X = np.array([["red"], ["blue"], ["green"], ["red"]])
enc = OneHotEncoder(sparse_output=False, handle_unknown="ignore")
enc.fit(X)
print(enc.transform([["red"], ["yellow"]]))
# [[0. 0. 1.]
#  [0. 0. 0.]]   <- unknown -> all zeros
print(enc.get_feature_names_out())   # ['x0_blue' 'x0_green' 'x0_red']

pandas get_dummies

import pandas as pd
df = pd.DataFrame({"color": ["red", "blue", "green", "red"]})
ohe = pd.get_dummies(df, columns=["color"], drop_first=True, dtype=int)
#    color_green  color_red
# 0            0          1
# 1            0          0
# 2            1          0
# 3            0          1

ColumnTransformer (production pipeline)

from sklearn.compose import ColumnTransformer
from sklearn.preprocessing import OneHotEncoder, StandardScaler
from sklearn.pipeline import Pipeline
from sklearn.linear_model import LogisticRegression

pre = ColumnTransformer([
    ("num", StandardScaler(), ["age", "income"]),
    ("cat", OneHotEncoder(handle_unknown="ignore"), ["city", "plan"]),
])
pipe = Pipeline([("pre", pre), ("clf", LogisticRegression(max_iter=1000))])
pipe.fit(X_train, y_train)

Sparse matrix 의 high-cardinality

enc = OneHotEncoder(sparse_output=True, handle_unknown="ignore")
X_sparse = enc.fit_transform(df[["zip_code"]])   # 40k columns sparse
# scipy.sparse.csr_matrix — memory-efficient

vs label encoding (decision)

from sklearn.preprocessing import LabelEncoder, OrdinalEncoder
# DON'T: feed LabelEncoder output to linear model
le = LabelEncoder()
y = le.fit_transform(["red", "blue", "green"])   # [2, 0, 1] — fake order!

# DO: OrdinalEncoder when order is real
oe = OrdinalEncoder(categories=[["low", "med", "high"]])

Frequency / target encoding (high-cardinality 대안)

import category_encoders as ce
te = ce.TargetEncoder(cols=["city"], smoothing=10)
X_tr = te.fit_transform(X_train, y_train)
X_te = te.transform(X_test)

Hashing trick (constant memory)

from sklearn.feature_extraction import FeatureHasher
h = FeatureHasher(n_features=256, input_type="string")
X_h = h.transform([["zip=" + z] for z in df["zip_code"]])

매 결정 기준

cardinality	model	encoding
<10	any	one-hot
10–50	linear / NN	one-hot or embedding
50–1000	tree	target / frequency
>1000	any	hashing / embedding
매 ordinal	any	OrdinalEncoder

기본값: OneHotEncoder(handle_unknown="ignore") in ColumnTransformer.

🔗 Graph

• 부모: Feature Engineering
• 변형: Target-Encoding
• 응용: Logistic-Regression-Foundations
• Adjacent: Sparse-Matrix · Curse-of-Dimensionality

🤖 LLM 활용

언제: 매 quick prototype, low-cardinality categorical, linear / tree baseline. 언제 X: 매 high-cardinality (>1000), text tokens (use embedding), online learning with new categories.

❌ 안티패턴

• Train-only fit: test set 의 unseen category 에 crash → handle_unknown="ignore".
• Drop-first with regularized model: 불필요한 정보 손실.
• OHE on high-cardinality without sparse: memory blowup.
• LabelEncoder for features: fake ordinal 강제, linear model 망가짐.
• Leak via target encoding without fold: target encoding 사용 시 K-fold 필수.

🧪 검증 / 중복

• Verified (sklearn 1.4 docs, pandas 2.2 docs).
• 신뢰도 A.

🕓 Changelog

날짜	변경
2026-05-08	Phase 1
2026-05-10	Manual cleanup — sklearn/pandas patterns + cardinality decision matrix