Isolation Forest のハイパーパラメータ調整

Anomaly Detection in Python

Bekhruz (Bex) Tuychiev

Kaggle Master, Data Science Content Creator

contamination の調整

決まった調整方法はない
依拠先:
- 直感
- EDA の知見
- ドメイン知識
- ビジネス要件

調査の例

類似の調査を調べる
最貧層と最富裕層の割合を把握
値を当てずっぽうで選ぶより調査が有効

Big Mart 売上データ

import pandas as pd

big_mart = pd.read_csv("big_mart_sales.csv")
big_mart.info()

<class 'pandas.core.frame.DataFrame'>
RangeIndex: 7060 entries, 0 to 7059
Data columns (total 5 columns):
 #   Column            Non-Null Count  Dtype  
 0   weight            7060 non-null   float64
 1   fat_content       7060 non-null   object 
 2   type              7060 non-null   object 
 3   max_retail_price  7060 non-null   float64
 4   sales             7060 non-null   float64
dtypes: float64(3), object(2)

カテゴリ変数のエンコード

big_mart = pd.get_dummies(big_mart)

   weight  max_retail_price      sales  fat_content_low_fat    fat_content_regular
0    9.30          249.8092  3735.1380                    1                      0
1    5.92           48.2692   443.4228                    0                      1
2   17.50          141.6180  2097.2700                    1                      0
3   19.20          182.0950   732.3800                    0                      1
4    8.93           53.8614   994.7052                    1                      0

evaluate_outlier_classifier

def evaluate_outlier_classifier(model, data):
    # ラベル取得
    labels = model.fit_predict(data)

    # インライヤーを返す
    return data[labels == 0]

evaluate_regressor

from sklearn.linear_model import LinearRegression
from sklearn.model_selection import train_test_split
from sklearn.metrics import root_mean_squared_error

evaluate_regressor

def evaluate_regressor(inliers):
    X = inliers.drop("sales", axis=1) 
    y = inliers[['sales']]


    X_train, X_test, y_train, y_test = train_test_split(X, y, random_state=10)

    lr = LinearRegression()
    lr.fit(X_train, y_train)


    preds = lr.predict(X_test)
    rmse = root_mean_squared_error(y_test, preds)

    return round(rmse, 3)

contamination の調整

contaminations = [0.05, 0.1, 0.2, 0.3]
scores = dict()


for c in contaminations:
    # 現在の c で IForest を生成
    iforest = IForest(contamination=c, random_state=10)


    # 現在の IForest でインライヤーを取得
    inliers = evaluate_outlier_classifier(iforest, big_mart)


    # RMSE を計算し scores に保存
    scores[c] = evaluate_regressor(inliers)

出力を確認

print(scores)

{0.05: 1148.555, 0.1: 1147.48, 0.2: 1082.307, 0.3: 1029.33}

複数ハイパーパラメータの調整

estimators = [100, 200, 300,]
max_samples = [0.6, 0.8, 1]
scores = dict()

デカルト積

from itertools import product

list(product(estimators, max_samples))

[(100, 0.6),
 (100, 0.8),
 (100, 1),
 (200, 0.6),
 (200, 0.8),
 (200, 1),
 (300, 0.6),
 (300, 0.8),
 (300, 1)]

ループ内の処理

estimators = [100, 200, 300,]
max_samples = [0.6, 0.8, 1]
scores = dict()


for e, m in product(estimators, max_samples):

    # IForest を生成
    iforest = IForest(n_estimators=e, max_samples=m, contamination=.3)


    # 現在の IForest でインライヤーを取得
    inliers = evaluate_outlier_classifier(iforest, big_mart)

    # RMSE を計算し scores に保存
    scores[(e, m)] = evaluate_regressor(inliers)

出力を確認

print(scores)

{(100, 0.6): 959.398,
 (100, 0.8): 986.056,
 (100, 1): 1195.875,
 (200, 0.6): 947.628,
 (200, 0.8): 933.115,
 (200, 1): 1195.875,
 (300, 0.6): 949.412,
 (300, 0.8): 935.962,
 (300, 1): 1195.875}

並列実行

# n_jobs=-1 で高速化
iforest = IForest(n_estimators=1000, n_jobs=-1)

iforest.fit(big_mart)

¡Vamos a practicar!

Anomaly Detection in Python