Cross validation 종류

K-fold Cross-Validation

---

 

Cross Validation 종류

 

1. K-fold Cross-validation

• 데이터셋을 K개의 sub-set으로 분리하는 방법
• 분리된 K개의 sub-set중 하나만 제외한 K-1개의 sub-sets를 training set으로 이용하여 K개의 모델 추정
• 일반적으로 K=5, K=10 사용 (-> 논문참고)
• K가 적어질수록 모델의 평가는 편중될 수 밖에 없음
• K가 높을수록 평가의 bias(편중된 정도)는 낮아지지만, 결과의 분산이 높을 수 있음

K-fold Cross Validation (K=5) 

2. Hold-out Validation

• Train/Test Split
• K를 1로 설정하여 하나의 학습/테스트 Split을 만들어 모델을 평가

3. LOOCV (Leave-one-out Cross-validation)

• fold 하나에 샘플 하나만 들어있는 K겹 교차 검증
• K를 전체 숫자로 설정하여 각 관측치가 데이터 세트에서 제외될 수 있도록 함
• 데이터셋이 클 때는 시간이 매우 오래 걸리지만, 작은 데이터셋에서는 좋은 결과를 만들어 냄
• 장점 : Data set에서 낭비 Data 없음
• 단점 : 측정 및 평가 고비용 소요

4. Repeated Random Sub Sampling Validation

• 임의로 Test set 선택
• 장점 : 측정 및 평가 저비용 소요
• 단점 : 미래 예측 시 신뢰성 예측 불가

5. Repeated K-fold Cross-validation

• k-fold cross-validation 과정을 n번 반복
• 반복할 때마다 데이터 샘플을 섞어서 샘플을 다르게 분할

6. Stratified K-fold Cross-validation

• 층화?
• 평균 응답 값이 모든 fold에서 대략 동일하도록 선택됨
• 각 fold가 전체를 잘 대표할 수 있도록 데이터를 재배열하는 프로세스

7. 중첩된 Cross-validation

• Cross-validation 내부에 Cross-validation 추가
• 쪼개고 쪼개기

8. Shuffle-split Cross-validation

• 임의 분할 교차 검증
• Stratified-Shuffle-Split 도 있음

Shuffle-split Cross-validation