Isolation Forest

1
2
3
4
5
6

import numpy as np
import pandas as pd
import seaborn as sns
from matplotlib import pyplot as plt
from sklearn.ensemble import IsolationForest
from sklearn.tree import plot_tree

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22

rs = np.random.RandomState(365)
dates = pd.date_range("1 1 2016", periods=365, freq="D")

data = {}
data["月"] = [d.strftime("%m") for d in dates]
data["曜日"] = [d.strftime("%A") for d in dates]
data["特徴1"] = [np.sin(d.day / 50) + np.random.rand() for d in dates]
data["特徴2"] = [np.cos(d.day / 50) + np.random.rand() for d in dates]
data["特徴3"] = [3 * np.random.rand() + np.log(d.dayofyear) * 0.03 for d in dates]
data["特徴4"] = [np.random.choice(["☀", "☂", "☁"]) for d in dates]
column_names = list(data.keys())

anomaly_index = [11, 49, 149, 240, 300, 310]
anomaly_dates = [dates[i] for i in anomaly_index]
for i in anomaly_index:
    data["特徴1"][i] = 2.5
    data["特徴2"][i] = 0.3

data = pd.DataFrame(data, index=dates)

plt.figure(figsize=(10, 4))
sns.lineplot(data=data, palette="tab10", linewidth=2.5)

1

data

1
2

X = pd.get_dummies(data)
X

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13

ilf = IsolationForest(
    n_estimators=100,
    max_samples="auto",
    contamination=0.01,
    max_features=5,
    bootstrap=False,
    random_state=np.random.RandomState(365),
)

ilf.fit(X)

data["is_anomaly"] = ilf.predict(X) < 0
data["anomaly_score"] = ilf.decision_function(X)

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11

plt.figure(figsize=(10, 4))
plt.title("検出箇所")
sns.lineplot(data=data[column_names], palette="tab10", linewidth=2.5)
for d in data[data["is_anomaly"]].index:
    plt.axvline(x=d, color="red", linewidth=4)

plt.figure(figsize=(10, 4))
plt.title("正解")
sns.lineplot(data=data[column_names], palette="tab10", linewidth=2.5)
for d in anomaly_dates:
    plt.axvline(x=d, color="black", linewidth=4)

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12

plt.figure(figsize=(24, 8))
plot_tree(
    ilf.estimators_[0],
    feature_names=column_names,
    filled=True,
    fontsize=13,
    max_depth=3,
    precision=2,
    rounded=True,
)

plt.show()