はじめてのFeature Engineering

import pandas as pd
from sklearn.preprocessing import LabelEncoder

df = pd.DataFrame({
    "category" : ["A","A","A","B","B","C"],
})

le = LabelEncoder()
df["encoded\_feature"] = le.fit_transform(df["category"])
df.head(10)

#one hot encoding sample
import pandas as pd
from sklearn.preprocessing import OneHotEncoder

df = pd.DataFrame({
    "category" : ["A","A","A","B","B","C"],
})

#文字をLabel Encodingして数値に変換
le = LabelEncoder()
df["encoded\_feature"] = le.fit_transform(df["category"])

#One Hot Encoding
ohe = OneHotEncoder()
ohe_array = ohe.fit_transform(df[["encoded\_feature"]])
df.drop("encoded\_feature", axis = 1, inplace = True)

#DataFrameに変換
ohe_df = pd.DataFrame(ohe_array.todense()).astype("uint8")
ohe_df.columns = ["category\_A","category\_B","category\_C"]

#元のDataFrameと結合
df = pd.concat([df,ohe_df], axis = 1)
df.head(10)

#Frequency Encoding Sample
import pandas as pd
from sklearn.preprocessing import OneHotEncoder

df = pd.DataFrame({
    "category" : ["A","A","A","B","B","C"],
})

#各カテゴリーの出現回数を計算
grouped = df.groupby("category").size().reset_index(name='category\_counts')

#元のデータセットにカテゴリーをキーとして結合
df = df.merge(grouped, how = "left", on = "category")
df["frequency"] = df["category\_counts"]/df["category\_counts"].count()

df.head(10)

#Target Mean Sample
import pandas as pd
from sklearn.preprocessing import OneHotEncoder

df = pd.DataFrame({
    "category" : ["A","A","A","B","B","C"],
    "outcome" : [1,1,0,1,0,1],
})

単純に計算すれば次のようなコードになる。

#各カテゴリーともう一つの特徴量の出現回数を計算
grouped_category = df.groupby("category")["category"].count().reset_index(name='category\_counts')
grouped_outcome = df.groupby("category")["outcome"].sum().reset_index(name='outcome\_counts')

#元のデータセットにカテゴリーをキーとして結合
df = df.merge(grouped_category, how = "left", on = "category")
df = df.merge(grouped_outcome, how = "left", on = "category")

df["target\_mean\_encoding"] = df["outcome\_counts"]/df["category\_counts"]

df.head(10)

#Target Mean Sample with leave one out scheme
import pandas as pd
from sklearn.preprocessing import OneHotEncoder

df = pd.DataFrame({
    "category" : ["A","A","A","B","B","C","C"],
    "outcome" : [1,1,0,1,0,1,1],
})

#各カテゴリーの出現回数を計算
grouped_category = df.groupby("category")["category"].count().reset_index(name='category\_counts')
grouped_outcome = df.groupby("category")["outcome"].sum().reset_index(name='outcome\_counts')

#元のデータセットにカテゴリーをキーとして結合
df = df.merge(grouped_category, how = "left", on = "category")
df = df.merge(grouped_outcome, how = "left", on = "category")

#計算する際に自分自身の値は除く
df["target\_mean\_encoding"] = (df["outcome\_counts"] - df["outcome"])/(df["category\_counts"] - 1)

df.head(10)