Skip to main content

特徵選擇_Deta_Selection_鐵人賽示範_

Open In Colab

17.特徵選擇

下載及整理資料​

import pandas as pd
import numpy as np

from sklearn.ensemble import RandomForestClassifier
from sklearn.feature_selection import RFE, SelectKBest, SelectFromModel, chi2, f_classif
from sklearn.model_selection import train_test_split
from sklearn.metrics import accuracy_score, roc_auc_score, precision_score, recall_score, f1_score
from sklearn.svm import LinearSVC
from sklearn.feature_selection import SelectFromModel
from sklearn.preprocessing import StandardScaler, MinMaxScaler

import seaborn as sns
import matplotlib
import matplotlib.pyplot as plt
# 下載鐵達尼號資料集
!wget -O data.csv https://raw.githubusercontent.com/duxuhao/Feature-Selection/master/example/titanic/clean_train.csv
df = pd.read_csv('data.csv')
df.describe(include='all')
# 確認有無缺失值
df.isna().sum()
# 特徵工程 (One-hot encoding)
df = pd.concat(
    [df,pd.get_dummies(df["Title"])], 
    axis=1
    )
df = df.drop("Title", 1)
# 切分特徵X與標籤Y
X = df.drop("Survived", 1)
Y = df["Survived"]
X.head()

定義評估模型​

  • 本範例採具有階層性的sklearn.ensemble.RandomForestClassifier訓練資料，評估結果為驗證資料集的metrics。
def use_RandomForestClassifier_evaluation_metrics_on_test_set(X,Y):
    X_train, X_test, Y_train, Y_test = train_test_split(
        X, Y, test_size = 0.2 ,stratify=Y, random_state = 9527)

    # 標準化
    scaler = StandardScaler().fit(X_train)
    X_train_scaled = scaler.transform(X_train)
    X_test_scaled = scaler.transform(X_test)

    # RandomForestClassifier訓練模型
    model = RandomForestClassifier(criterion='entropy', random_state=9527)
    model.fit(X_train_scaled, Y_train)

    # 預測
    y_predict_result = model.predict(X_test_scaled)

    # 回傳evaluation_metrics_on_test_set
    return {
        'accuracy' : accuracy_score(Y_test, y_predict_result),
        'roc' : roc_auc_score(Y_test, y_predict_result),
        'precision' : precision_score(Y_test, y_predict_result),
        'recall' : recall_score(Y_test, y_predict_result),
        'f1' : f1_score(Y_test, y_predict_result),
        'Feature Count' : len(X.columns)
        }

全特徵原始成效​

res = pd.DataFrame(use_RandomForestClassifier_evaluation_metrics_on_test_set(X,Y), index=['ALL'])
res
# Correlation Matrix
plt.figure(figsize=(15,15))
cor = df.corr() 
sns.heatmap(cor, annot=True)
plt.show()

過濾方法 Filter Method​

依關聯性移除特徵​

# 取得具有與其他部分特徵高度相關的某特徵絕對值
cor_target = abs(cor["FamilySize"])

# 選擇高度相關的特徵（閾值 = 0.2）
relevant_features = cor_target[cor_target>0.2]

# 選擇特徵名稱
names = [index for index, value in relevant_features.iteritems()]

# 刪除目標特徵
names.remove('FamilySize')

print(names)
res = res.append(
    pd.DataFrame(
        use_RandomForestClassifier_evaluation_metrics_on_test_set(
            X[names],
            Y), 
            index=['Remove High Corr']))
res
# Correlation Matrix
plt.figure(figsize=(15,15))
cor = X[names].corr() 
sns.heatmap(cor, annot=True,)
plt.show()

單變量特徵選取 Univariate Selection​

def univariate_selection(X, Y, k=10):
    
    X_train, X_test, Y_train, Y_test = train_test_split(X, Y, test_size = 0.2, stratify=Y, random_state = 9527)
    
    scaler = StandardScaler().fit(X_train)
    X_train_scaled = scaler.transform(X_train)
    X_test_scaled = scaler.transform(X_test)
    
    # User SelectKBest to select top 10 features based on f-test
    selector = SelectKBest(f_classif)
    X_new = selector.fit_transform(X_train_scaled, Y_train)

    feature_idx = selector.get_support()

    feature_names = X.columns[feature_idx]
    
    return feature_names
univariate_selection(X,Y)
res = res.append(
    pd.DataFrame(
        use_RandomForestClassifier_evaluation_metrics_on_test_set(
            X[univariate_selection(X,Y)],
            Y), 
            index=['Univariate Selection']))
res

包裝方法 Wrapper Method​

遞迴特徵消除 Recursive feature elimination (RFE)​

# Recursive Feature Elimination
def rfe_selection( X , Y, k=10):
    
    # Split train and test sets
    X_train, X_test, Y_train, Y_test = train_test_split(
        X, 
        Y, 
        test_size = 0.2, 
        stratify=Y, 
        random_state = 9527)
    
    scaler = StandardScaler().fit(X_train)
    X_train_scaled = scaler.transform(X_train)
    X_test_scaled = scaler.transform(X_test)
    
    model = RandomForestClassifier(
        criterion='entropy', 
        random_state=9527
        )
    rfe = RFE(model)
    rfe = rfe.fit(X_train_scaled, Y_train)

    feature_names = X.columns[rfe.get_support()]
    
    return feature_names

rfe_selection(X,Y,10)
res = res.append(
    pd.DataFrame(
        use_RandomForestClassifier_evaluation_metrics_on_test_set(
            X[rfe_selection(X,Y)],
            Y), 
            index=['RFE']))

res

嵌入方法 Embedded Method​

重要特徵 Feature importance​

def feature_importance(X,Y):
    X_train, X_test, Y_train, Y_test = train_test_split(X, Y, test_size = 0.2,stratify=Y, random_state = 9527)
    
    scaler = StandardScaler().fit(X_train)
    X_train_scaled = scaler.transform(X_train)
    X_test_scaled = scaler.transform(X_test)

    model = RandomForestClassifier()
    model = model.fit(X_train_scaled,Y_train)

    plt.figure(figsize=(10, 12))
    feat_importances = pd.Series(model.feature_importances_, index=X.columns)
    feat_importances.sort_values().plot(kind='barh')

    plt.show()
    return model


def select_features_from_model(model,X):
    
    model = SelectFromModel(model, prefit=True, threshold=0.013)
    feature_idx = model.get_support()
    feature_names = X.columns[feature_idx]
        
    return feature_names

model = feature_importance(X,Y)
feature_imp_feature_names = select_features_from_model(model,X)
feature_imp_feature_names
res = res.append(
    pd.DataFrame(
        use_RandomForestClassifier_evaluation_metrics_on_test_set(
            X[feature_imp_feature_names],
            Y), 
            index=['Feature Importance']))
res

L1正規化 L1 regularization​

def run_l1_regularization(X,Y):
    
    # Split train and test set
    X_train, X_test, Y_train, Y_test = train_test_split(X, Y, test_size = 0.2,stratify=Y, random_state = 123)
    
    # All features of dataset are float values. You normalize all features of the train and test dataset here.
    scaler = StandardScaler().fit(X_train)
    X_train_scaled = scaler.transform(X_train)
    X_test_scaled = scaler.transform(X_test)
    
    # Select L1 regulated features from LinearSVC output 
    selection = SelectFromModel(LinearSVC(C=1, penalty='l1', dual=False))
    selection.fit(X_train_scaled, Y_train)

    feature_names = X.columns[(selection.get_support())]
    
    return feature_names

l1reg_feature_names = run_l1_regularization(X,Y)
res = res.append(
    pd.DataFrame(
        use_RandomForestClassifier_evaluation_metrics_on_test_set(
            X[run_l1_regularization(X,Y)],
            Y), 
            index=['L1']))
res

評估小結​

final_res = pd.DataFrame(use_RandomForestClassifier_evaluation_metrics_on_test_set(X,Y), index=['ALL'])
final_res = final_res.append(pd.DataFrame(use_RandomForestClassifier_evaluation_metrics_on_test_set(X[names],Y), index=['Remove High Corr']))
final_res = final_res.append(pd.DataFrame(use_RandomForestClassifier_evaluation_metrics_on_test_set(X[univariate_selection(X,Y)],Y), index=['Univariate Selection']))
final_res = final_res.append(pd.DataFrame(use_RandomForestClassifier_evaluation_metrics_on_test_set(X[rfe_selection(X,Y)],Y), index=['RFE']))
final_res = final_res.append(pd.DataFrame(use_RandomForestClassifier_evaluation_metrics_on_test_set(X[run_l1_regularization(X,Y)],Y), index=['L1']))
final_res.sort_values('accuracy', ascending=False)

參考​