特徵工程與選擇

端到端機器學習

Joshua Stapleton

Machine Learning Engineer

特徵工程

建立特徵

簡化問題
提升模型效率

技巧

修改現有特徵
設計新特徵

好處

更易部署、維護、訓練
提升可解釋性

機器學習生命週期目前階段：特徵工程

正規化（Normalization）

將數值特徵縮放到 [0, 1]
當特徵尺度／範圍不一致時很有幫助。

from sklearn.model_selection import train_test_split
from sklearn.preprocessing import Normalizer

# Split the data
X_train, X_test = train_test_split(df, test_size=0.2, random_state=42)
# Createnormalizer object, fit on training data, normalize, and transform test set
norm = Normalizer()
X_train_norm = norm.fit_transform(X_train)
X_test_norm = norm.transform(X_test)

標準化（Standardization）

將資料縮放為 mean = 0、variance = 1
對假設平均與變異相近的演算法有利

from sklearn.preprocessing import StandardScaler

# Split the data
X_train, X_test = train_test_split(df, test_size=0.2, random_state=42)
# Create a scaler object and fit training data to standardize it
sc = StandardScaler()
X_train_stzd = sc.fit_transform(X_train)
# Only standardize the test data
X_test_stzd = sc.transform(X_test)

好的特徵要素是什麼？

使用相關特徵
看診當天的天氣不應影響診斷

暴風雨天氣的圖片，說明特徵選擇中的相關性原則

使用不相似（正交）的特徵
「以月為單位的年齡」與「以年為單位的年齡」同時存在並無助益

示意圖：特徵選擇中的正交性原則

sklearn.feature_selection

from sklearn.ensemble import RandomForestClassifier
from sklearn.feature_selection import SelectFromModel
from sklearn.model_selection import train_test_split


# Splitting data into train and test subsets first to avoid data leakage
X_train, X_test, y_train, y_test = train_test_split(
    heart_disease_df_X, heart_disease_df_y, test_size=0.2, random_state=42)

sklearn.feature_selection（續）

# Define and fit the random forest model
rf = RandomForestClassifier(n_jobs=-1, class_weight='balanced', max_depth=5)
rf.fit(X_train, y_train)


# Define and run feature selection
model = SelectFromModel(rf, prefit=True)
features_bool = model.get_support()
features = heart_disease_df.columns[features_bool]

一起來練習吧！

端到端機器學習