异常值分类器集成

Python 中的异常检测

Bekhruz (Bex) Tuychiev

Kaggle Master, Data Science Content Creator

回到 Airbnb

outliers = airbnb_df[iforest.labels_ == 1]

outlier_probs = iforest.predict_proba(outliers)
outlier_probs[:10]

array([[0.51999538, 0.48000462],
       [0.61789522, 0.38210478],
       [0.61802032, 0.38197968],
       [0.35184434, 0.64815566],
       [0.57533286, 0.42466714],
       [0.59038933, 0.40961067],
       [0.57677613, 0.42322387],
       [0.54158826, 0.45841174],
       [0.49118093, 0.50881907],
       [0.21387357, 0.78612643]])

概率阈值最佳实践

阈值：

低风险用 75%
高成本场景用 >90%，如：
- 医疗
- 网络安全
- 数据有限

什么是集成？

两个或以上分类器的组合
预测更稳定

查看数据

google.head()

              Open    High     Low   Close    Volume  day_of_week  month  day
Date                                                                         
2006-01-03  211.47  218.05  209.32  217.83  13137450            1      1    3
2006-01-04  222.17  224.70  220.09  222.84  15292353            2      1    4
2006-01-05  223.22  226.00  220.97  225.85  10815661            3      1    5
2006-01-06  228.66  235.49  226.85  233.06  17759521            4      1    6
2006-01-09  233.44  236.94  230.70  233.68  12795837            0      1    9

缩放数值特征

from sklearn.preprocessing import QuantileTransformer


# Define the cols to be scaled
to_scale = ['Open', 'High', 'Low', 'Close', 'Volume']


# Initiate the transformer
qt = QuantileTransformer(output_distribution="normal")

# Scale and store the columns back
google.loc[:, to_scale] = qt.fit_transform(google[to_scale])

创建数组

# Create a list of estimators
estimators = [KNN(n_neighbors=20), LOF(n_neighbors=20), IForest()]


# Create an empty array
shape = (len(google), len(estimators))
probability_scores = np.empty(shape=shape)

循环内部

estimators = [KNN(n_neighbors=20), LOF(n_neighbors=20), IForest()]

shape = (len(google), len(estimators))
probability_scores = np.empty(shape=shape)


# Loop over and fit
for index, est in enumerate(estimators):
    est.fit(google)


    # Create probabilities
    probs = est.predict_proba(google)


    # Store the probs
    probability_scores[:, index] = probs[:, 1]

聚合：均值

mean_scores = np.mean(probability_scores, axis=1)

mean_scores

array([0.20699869, 0.21455413, 0.17166271, ..., 0.31255075, 0.33553513,
       0.32217186])

聚合：中位数

median_scores = np.mean(probability_scores, axis=1)

median_scores

array([0.20699869, 0.21455413, 0.17166271, ..., 0.31255075, 0.33553513,
       0.32217186])

概率筛选

# Create a mask with 75% threshold
is_outlier = median_scores > 0.75


# Filter the outliers
outliers = google[is_outlier]


len(outliers)

步骤汇总

# Create a list of estimators
estimators = [KNN(n_neighbors=20), LOF(n_neighbors=20), IForest()]
probability_scores = np.empty(shape=(len(google), len(estimators)))


for index, est in enumerate(estimators):
    # Fit and generate probabilities
    est.fit(google)
    probs = est.predict_proba(google)

    # Store the probabilities
    probability_scores[:, index] = probs[:, 1]

步骤汇总

# Average the scores
mean_scores = np.mean(probability_scores, axis=1)

# Filter with 75% threshold
outliers = google[mean_scores > 0.75]

print(len(outliers))

Vamos praticar!

Python 中的异常检测