anomaly-detector

异常检测器

适用对象： 检测数据集异常值的数据工程师和分析师。

目标： 为各种数据类型提供可用于生产环境的异常检测函数。

脚本

从 scripts/anomaly_detection.py 执行检测函数：

from scripts.anomaly_detection import (
    detect_anomalies_zscore,
    detect_anomalies_iqr,
    detect_anomalies_modified_zscore,
    detect_anomalies_isolation_forest,
    detect_anomalies_lof,
    detect_anomalies_rolling,
    detect_anomalies_stl,
    detect_anomalies_ensemble
)

方法选择

方法	最适用场景	局限性
Z-Score	正态分布	对异常值敏感
IQR	偏态分布	整体灵敏度较低
Modified Z-Score	稳健检测	计算速度较慢
Isolation Forest	高维数据	需要调参
LOF	局部密度异常	计算开销大
Rolling	存在趋势的时间序列	对窗口大小敏感
STL	季节性时间序列	需要已知周期

使用示例

单列检测

import pandas as pd
from scripts.anomaly_detection import detect_anomalies_zscore, detect_anomalies_iqr

df = pd.read_csv('data.csv')

# Z-score method (good for normal distributions)
anomalies_z = detect_anomalies_zscore(df['value'], threshold=3.0)

# IQR method (robust to skewed data)
anomalies_iqr = detect_anomalies_iqr(df['value'], multiplier=1.5)

print(f"Z-score found {anomalies_z.sum()} anomalies")
print(f"IQR found {anomalies_iqr.sum()} anomalies")

使用 Isolation Forest 进行多列检测

from scripts.anomaly_detection import detect_anomalies_isolation_forest

numeric_cols = ['revenue', 'quantity', 'price']
anomalies = detect_anomalies_isolation_forest(df, numeric_cols, contamination=0.01)

df_anomalies = df[anomalies]

集成方法（推荐）

from scripts.anomaly_detection import detect_anomalies_ensemble

results = detect_anomalies_ensemble(
    df,
    columns=['revenue', 'quantity'],
    methods=['zscore', 'iqr', 'isolation_forest'],
    min_agreement=2  # Flag if 2+ methods agree
)

confirmed_anomalies = df[results['is_anomaly']]

时间序列异常

from scripts.anomaly_detection import detect_anomalies_rolling, detect_anomalies_stl

# Rolling window (for trending data)
anomalies = detect_anomalies_rolling(df['daily_sales'], window=7, n_std=2.0)

# STL decomposition (for seasonal data)
anomalies = detect_anomalies_stl(df['monthly_revenue'], period=12, threshold=3.0)

依赖项

pandas
numpy
scikit-learn  # For Isolation Forest, LOF
statsmodels   # For STL decomposition