04단계 · 2유형 — 예측과 제출
데이터 누수 — 평가 데이터에 fit 하지 않는다
전처리 기준을 전체로 배울 때와 학습용으로만 배울 때의 차이를 숫자로 확인한다.
먼저 알아볼게요
코드로 이동 ↓데이터 누수(leakage) 는 모델이 "시험 문제를 미리 본" 상태를 말한다. 그러면 검증 점수가 실제 실력보다 좋게 나온다 — 내가 내 점수를 못 믿게 되는 것이 문제다. 실제 성능이 반드시 나빠진다는 뜻은 아니지만, 좋은지 나쁜지를 알 수 없게 된다.
자주 만나는 누수는 전처리를 전체 데이터로 fit 하는 것이다. 평균·표준편차·최솟값·최댓값을 구할 때 평가 데이터가 섞이면, 그 정보가 이미 모델에 들어간 것이다.
규칙은 하나다: `fit` 은 학습 데이터에만, `transform` 은 양쪽에. 그래서 학습용은 fit_transform, 검증·평가용은 transform 을 쓴다.
scaler.mean_ 에 fit 할 때 배운 평균이 들어 있다. 전체로 fit 한 것과 학습용으로만 fit 한 것의 mean_ 을 나란히 찍어 보면 값이 다르다 — 누수는 추상적인 말이 아니라 이 숫자 차이다.
결측치를 채울 때도 같다. 채울 값(평균·중앙값)은 학습 데이터에서 구해 평가 데이터에도 그 값을 쓴다.
이렇게 써요
import pandas as pd
from sklearn.preprocessing import StandardScaler
from sklearn.model_selection import train_test_split
train = pd.read_csv("churn_train.csv")
X = train[["가입개월", "월요금"]]
x_tr, x_va = train_test_split(X, test_size=0.2, random_state=42)
sc = StandardScaler().fit(x_tr)
print(sc.mean_.round(4).tolist())기억할 한 가지
fit 은 학습 데이터에만, transform 은 양쪽에. 결측치 채울 값도 학습 데이터에서 구한다.
직접 해보기
churn_train.csv 의 가입개월·월요금·부가서비스 로 ① 전체에 fit 한 스케일러의 월요금 평균(소수 4) ② 학습용에만 fit 한 스케일러의 월요금 평균(소수 4) ③ 두 스케일러로 바꾼 검증 데이터의 최대 차이(소수 6) ④ 그 차이가 0 보다 큰지를 출력하세요.
제공 파일 1개 · 내용 보기
파일은 준비되어 있어요. 코드에서 이름으로 불러오세요.
churn_train.csv
고객ID,성별,요금제,가입개월,월요금,부가서비스,이탈 C1001,남,standard,22,44300,3,1 C1002,여,premium,1,62000,1,1 C1003,남,premium,45,64000,0,1 C1004,남,basic,10,32200,1,1 C1005,남,premium,5,69500,0,1 C1006,여,basic,56,29500,0,0 C1007,여,premium,48,62800,2,0 C1008,여,standard,5,56700,1,1 C1009,여,basic,49,30600,3,0 C1010,여,premium,52,74800,2,1 C1011,여,basic,19,32800,2,1 C1012,여,standard,18,42200,3,0 C1013,남,standard,37,47900,1,0 C1014,남,basic,1,30500,4,1 C1015,여,standard,52,43200,0,0 C1016,남,premium,31,75300,3,1 C1017,여,basic,22,29100,1,1 C1018,남,standard,39,48400,1,1 C1019,남,premium,35,82500,0,0 C1020,남,standard,19,42600,3,0 C1021,남,basic,22,27300,4,1 C1022,여,basic,38,27800,1,1 C1023,남,standard,5,54200,0,1 C1024,여,premium,36,77200,3,0 C1025,여,standard,14,46500,1,1 C1026,남,premium,15,64500,2,1 C1027,남,basic,45,28100,4,0 C1028,남,standard,49,40600,1,1 C1029,남,standard,59,50800,4,0 C1030,여,basic,16,27800,3,1 C1031,남,basic,36,24900,0,1 C1032,남,basic,17,30200,4,0 C1033,여,basic,36,29900,3,1 C1034,남,basic,36,23800,2,0 C1035,여,standard,9,50200,1,1 C1036,여,premium,45,60800,2,0 C1037,여,standard,37,42900,3,0 C1038,여,basic,37,24800,2,0 C1039,남,standard,42,46300,0,0 C1040,여,basic,34,28400,2,1 C1041,남,premium,11,84200,3,1 C1042,남,standard,9,57000,4,1 C1043,여,standard,11,43500,3,1 C1044,여,basic,42,29200,2,1 C1045,여,premium,10,66800,4,1 C1046,남,premium,56,62700,2,0 C1047,남,standard,19,44100,3,1 C1048,여,basic,14,33700,2,0 C1049,여,basic,44,28800,0,1 C1050,여,premium,32,69800,0,1 C1051,남,basic,52,24100,3,0 C1052,남,standard,48,35900,1,0 C1053,여,basic,41,26300,0,0 C1054,여,basic,36,25500,1,1 C1055,여,basic,7,29000,3,0 C1056,여,basic,34,32700,3,1 C1057,남,standard,42,48900,4,1 C1058,여,basic,32,25800,1,1 C1059,남,standard,56,50100,3,0 C1060,여,standard,28,41500,1,0 C1061,여,premium,47,73900,0,0 C1062,남,premium,4,53900,3,1 C1063,여,premium,4,72600,0,1 C1064,여,basic,42,29100,4,1 C1065,남,basic,59,29800,4,0 C1066,남,premium,28,75100,1,0 C1067,남,basic,27,27200,0,0 C1068,남,basic,60,30100,2,1 C1069,여,premium,4,67400,1,1 C1070,여,basic,4,34100,0,1 C1071,남,standard,42,40800,4,0 C1072,여,premium,53,72900,4,0 C1073,남,standard,36,37500,1,0 C1074,남,standard,44,49800,0,0 C1075,남,basic,5,29700,3,1 C1076,남,premium,1,75600,3,1 C1077,남,standard,44,40600,1,1 C1078,여,basic,7,23800,1,0 C1079,여,premium,47,66100,4,1 C1080,남,standard,32,41500,1,0 C1081,남,premium,37,78500,4,1 C1082,남,premium,52,68500,1,0 C1083,남,premium,51,61900,3,1 C1084,남,basic,38,30000,2,1 C1085,남,basic,12,30900,0,0 C1086,남,premium,25,74900,1,1 C1087,여,standard,40,47300,1,0 C1088,남,standard,6,41600,1,1 C1089,남,premium,57,64200,4,0 C1090,여,basic,42,28700,0,1 C1091,여,standard,27,37300,0,1 C1092,남,basic,27,25900,2,1 C1093,여,basic,41,27900,0,0 C1094,남,standard,60,44200,0,0 C1095,남,premium,12,67100,2,1 C1096,남,standard,46,45700,1,0 C1097,여,standard,23,48800,1,1 C1098,남,premium,57,75300,4,0 C1099,남,premium,13,53900,0,0 C1100,여,premium,29,61600,3,1 C1101,남,standard,2,47300,3,1 C1102,남,standard,10,55900,1,1 C1103,남,premium,3,71800,1,1 C1104,남,premium,53,70000,1,0 C1105,남,premium,32,71300,2,1 C1106,남,premium,9,66000,0,1 C1107,여,premium,30,66100,2,1 C1108,남,standard,43,53500,3,0 C1109,남,basic,26,29400,4,0 C1110,여,standard,21,48500,4,0 C1111,남,standard,51,42400,0,1 C1112,남,premium,21,74100,4,0 C1113,남,premium,56,56600,0,0 C1114,남,standard,46,45300,4,0 C1115,여,standard,48,42300,4,0 C1116,여,basic,35,24700,4,1 C1117,여,premium,22,42700,1,0 C1118,남,standard,9,39900,4,1 C1119,남,premium,36,73100,1,0 C1120,여,premium,38,73800,0,1 C1121,여,standard,13,49300,4,0 C1122,남,basic,48,30300,0,0 C1123,남,standard,16,43800,0,1 C1124,남,premium,15,74000,2,1 C1125,남,basic,25,30200,0,1 C1126,남,basic,44,33000,2,1 C1127,남,basic,56,21800,3,0 C1128,남,standard,48,38600,1,0 C1129,여,premium,15,68200,0,1 C1130,여,basic,28,27400,3,0 C1131,남,standard,54,40800,4,0 C1132,남,basic,45,26900,3,0 C1133,여,premium,21,66200,0,1 C1134,여,premium,39,58900,0,0 C1135,남,basic,43,29000,3,0 C1136,남,standard,9,48600,4,1 C1137,남,basic,28,30300,1,0 C1138,남,premium,30,78600,4,0 C1139,남,basic,18,23900,3,0 C1140,남,standard,46,40900,2,1 C1141,남,premium,6,70300,0,1 C1142,남,basic,46,29100,1,0 C1143,남,standard,18,41600,3,0 C1144,여,basic,53,27800,3,1 C1145,여,basic,7,25400,0,1 C1146,남,basic,27,28800,0,1 C1147,여,premium,51,64300,4,0 C1148,남,premium,25,53100,2,0 C1149,남,standard,54,43800,1,1 C1150,남,premium,40,57700,0,1 C1151,여,standard,24,42900,3,1 C1152,남,premium,21,63300,1,1 C1153,여,standard,44,34800,3,0 C1154,여,basic,29,27400,2,1 C1155,여,premium,38,61500,1,0 C1156,여,premium,48,62600,4,0 C1157,남,basic,27,26900,1,1 C1158,남,standard,20,38000,4,0 C1159,여,premium,22,63700,2,1 C1160,여,standard,20,42800,1,0
내 코드 Python
Python 준비 중… 처음 한 번만 내려받아요
들여쓰기: Tab · 편집기에서 나가기: Esc 다음 Tab
실행 결과
실행하면 코드가 출력한 내용이 여기에 나타나요.
막혔을 때
풀이와 비교하기
내 코드와 한 줄씩 비교해 보세요. 풀이를 보는 것만으로 완료되지는 않아요.
import numpy as np
import pandas as pd
from sklearn.model_selection import train_test_split
from sklearn.preprocessing import StandardScaler
train = pd.read_csv("churn_train.csv")
X = train[["가입개월", "월요금", "부가서비스"]]
y = train["이탈"]
x_tr, x_va, y_tr, y_va = train_test_split(X, y, test_size=0.2, random_state=42)
wrong = StandardScaler().fit(X)
right = StandardScaler().fit(x_tr)
print(round(float(wrong.mean_[1]), 4))
print(round(float(right.mean_[1]), 4))
gap = float(np.abs(wrong.transform(x_va) - right.transform(x_va)).max())
print(round(gap, 6))
print(gap > 0)