def average_KFold(X, y, fold, start_deg, finish_deg, error_type):
kf = KFold(n_splits=fold)
error_test, error_train = [], []
KFold(n_splits=5, random_state=None, shuffle=False)
for i in range(start_deg, finish_deg + 1):
err1, err2 = 0, 0
for train_index, test_index in kf.split(X):
X_train, X_test = X[train_index], X[test_index]
y_train, y_test = y[train_index], y[test_index]
P = poly(i)
P_train = P.fit_transform(X_train)
a = np.linalg.inv(P_train.transpose().dot(P_train))
w = a.dot(P_train.transpose())
W = w.dot(y_train)
P_test = P.fit_transform(X_test)
P_train = P.fit_transform(X_train)
prediction_test = P_test.dot(W)
prediction_train = P_train.dot(W)
err1 += MSE(y_test, prediction_test)
err2 += MSE(y_train, prediction_train)
err1 = err1 / fold
err2 = err2 / fold
error_test.append(err1)
error_train.append(err2)
return error_test, error_train
def regr_with_reg(X_train, y_train, to_predict, deg, a):
rr = Ridge(alpha=a)
p = poly(degree=deg)
X_train_poly = p.fit_transform(X_train)
to_predict_poly = p.fit_transform(to_predict)
rr.fit(X_train_poly, y_train)
y_pred = rr.predict(to_predict_poly)
return y_pred
def regr_without_reg(X_train, y_train, X_test, deg):
p = poly(degree=deg)
X_train_poly = p.fit_transform(X_train)
model = LinearRegression()
model.fit(X_train_poly, y_train)
X_test_poly = p.fit_transform(X_test)
y_pred = model.predict(X_test_poly)
return y_pred
def makeInput(data):
x = pd.DataFrame()
x["agena"] = data.Age.map(lambda x:1 if math.isnan(x) else 0)
x["farena"] = data.Fare.map(lambda x:1 if math.isnan(x) else 0)
data.Age = data.Age.fillna(29.7)
data.Fare = data.Fare.fillna(32.2)
data.Cabin = data.Cabin.fillna("NA")
x["cabinna"] = data.Cabin.map(lambda x:1 if x=="NA" else 0)
x["sibsp"] = data.SibSp
x["parch"] = data.Parch
x["smallfamiliy"] = (data.SibSp+data.Parch).map(lambda x:1 if x<3 else 0)
x["bigfamiliy"] = (data.SibSp+data.Parch).map(lambda x:1 if x>=3 else 0)
x["age"] = data.Age.map(lambda x: (x-29.7)/13.0)
x["age-10"] = data.Age.map(lambda x:1 if x<=10 else 0)
x["age10-15"] = data.Age.map(lambda x:1 if x>10 and x<=15 else 0)
x["age15-20"] = data.Age.map(lambda x:1 if x>15 and x<=20 else 0)
x["age20-25"] = data.Age.map(lambda x:1 if x>20 and x<=25 else 0)
x["age25-30"] = data.Age.map(lambda x:1 if x>25 and x<31 else 0)
x["age30-"] = data.Age.map(lambda x:1 if x>30 else 0)
x["class1"] = data.Pclass.map(lambda x:1 if x==1 else 0)
x["class2"] = data.Pclass.map(lambda x:1 if x==2 else 0)
x["class3"] = data.Pclass.map(lambda x:1 if x==3 else 0)
x["male"] = data.Sex.map(lambda x:1 if x=="male" else 0)
x["female"] = data.Sex.map(lambda x:1 if x=="female" else 0)
x["fare"] = data.Fare.map(lambda x: (x-32.2)/49.7)
x["fare-"] = data.Fare.map(lambda x:1 if x<20 else 0)
x["fare+"] = data.Fare.map(lambda x:1 if x>=20 else 0)
x["mrs"] = data.Name.map(lambda x:1 if x.lower().find("mrs")>=0 else 0)
x["mr"] = data.Name.map(lambda x:1 if x.lower().find("mr")>=0 else 0)
x["miss"] = data.Name.map(lambda x:1 if x.lower().find("miss")>=0 else 0)
x["master"] = data.Name.map(lambda x:1 if x.lower().find("master")>=0 else 0)
x["embark_C"] = data.Embarked.map(lambda x:1 if x=="C" else 0)
x["embark_Q"] = data.Embarked.map(lambda x:1 if x=="Q" else 0)
x["embark_S"] = data.Embarked.map(lambda x:1 if x=="S" else 0)
#return x
p = poly(2, interaction_only=False)
return p.fit_transform(x)
stu_data.iloc[:,:-2] = f_dummies(stu_data.iloc[:,:-2])
# 2) scaling
m_sc = standard()
m_sc.fit(stu_data)
stu_x_sc = m_sc.transform(stu_data)
# 4. data 분리
train_x, test_x, train_y, test_y = train_test_split(stu_x_sc,
stu_target,
random_state=0)
# 1. interaction 적용 data 추출
from sklearn.preprocessing import PolynomialFeatures as poly
m_poly = poly(degree=2)
m_poly.fit(train_x) # 각 설명변수에 대한 2차항 모델 생성
train_x_poly = m_poly.transform(train_x) # 각 설명변수에 대한 2차항 모델 생성
test_x_poly = m_poly.transform(test_x)
m_poly.get_feature_names() # 변경된 설명변수들의 형태 확인
col_poly = m_poly.get_feature_names(stu_data.columns) # 실제 컬럼이름의 교호작용 출력
DataFrame(m_poly.transform(train_x) ,
columns = m_poly.get_feature_names(stu_data.columns))
# 2. 확장된 데이터셋을 RF에 학습, feature importance 확인
m_rf = rf(random_state=0)
m_rf.fit(train_x_poly, train_y)
m_rf.score(test_x_poly, test_y) # 0.71717
# --------------------------------------------------------------------------- #
# [ 분석 시 고려사항 3. 교호작용 ]
# - 변수 상호 간 서로 결합된 형태로 의미 있는 경우
# - 2차, 3차항 ... 추가 가능
# - 발생 가능한 모든 다차항의 interaction으로 부터 의미 있는 변수 추출
# 3.1 interaction 적용 data 추출
from sklearn.preprocessing import PolynomialFeatures as poly
원본 => 2차항 적용 (transform 작업)
x1 x2 x3 x1^2 x2^2 x3^2 x1x2 x1x2 x2x3
1 2 3 1 4 9 2 3 6
2 4 5 4 16 25 8 10 20
m_poly = poly(degree = 2) # 2차항을 만들겠다
m_poly.fit(train_x) # 각 설명변수에 2차항 모델 생성
# ** test data set은 fitting 필요 없음 why?
train_x_poly = m_poly.transform(train_x) # 스케일링 된 데이터셋으로 하는게 더 좋음
test_x_poly = m_poly.transform(test_x)
m_poly.get_feature_names() # 변경된 설명변수들의 형태 = 2차항 모습
DataFrame(m_poly.transform(train_x),
columns = m_poly.get_feature_names()) # 보기 좋음 -> 변수가 엄청 많으면 이 또한 쉽진 않음
col_poly = m_poly.get_feature_names(df_iris.feature_names) # 실제 컬럼이름이 반영된 교호작용 출력
DataFrame(m_poly.transform(train_x),
columns = m_poly.get_feature_names(df_iris.feature_names)) # 훨씬 보기 좋음
from sklearn.preprocessing import MinMaxScaler as minmax m_sc2 = minmax() m_sc2.fit(train_x) train_x_sc2 = m_sc2.transform(train_x) m_sc2.fit(test_x) test_x_sc2 = m_sc2.transform(test_x) # step 2) 데이터 학습 & 평가 m_knn.fit(train_x_sc2, train_y) m_knn.score(test_x_sc2, test_y) # 0.959 => min_max 경우 scaling 전보다 상승* # 3) 전체 interaction 학습 (min_max로 scaling된 상태) # step 1) 모델 생성 from sklearn.preprocessing import PolynomialFeatures as poly m_poly = poly(degree=2) # 2차항까지 m_poly.fit(train_x_sc2) train_x_poly = m_poly.transform(train_x_sc2) test_x_poly = m_poly.transform(test_x_sc2) col_poly = m_poly.get_feature_names(df_cancer.feature_names) # step 2) 데이터 학습 & 평가 m_knn.fit(train_x_poly, train_y) m_knn.score(test_x_poly, test_y) # 0.964 => 2차항까지 교호작용한 설명변수로 이전 값보다 상승 ** # 4) 선택된 interaction 학습 # 변형된 데이터셋 Rf에 학습 후 feature importance 확인 m_rf = rf_c(random_state=0) m_rf.fit(train_x_poly, train_y)