[[2021__CSUR__A Review on Outlier-Anomaly Detection in Time Series Data]] [Chapter 5 Outlier detection in Time series | Time Series with R](https://s-ai-f.github.io/Time-Series/outlier-detection-in-time-series.html) ## 入力 - 単変量時系列 - 多変量時系列 ## 出力タイプ - Point outlier - Subsequences - Time Series - 多変量のみ ## 検出手法 - Model-based - Estimation: - Median Absolute Deviation (MAD) - Exponentially Weighted Moving Average (EWMA) method - Extreme Studentized Deviate (ESD) - STL decomposition - Prediction: - ARIMA model - ARIMA model within a sliding window to compute the prediction interval, so the parameters are refitted each time that the window moves a step forward. - Extreme value theory - Density-based - [[k近傍法による時系列データの異常部位検出]] - Histogramming ## 時系列データの異常の種類 - Additive outliers - Temporal changes - Level shifts ## 方法論的アプローチ - Statistical based methods - Forecasting-based approaches - Neural Network Based Approaches - Clustering Based Approaches - Proximity Based Approaches - Tree Based Approaches - Dimension Reduction Based Approaches