Anomaly Detection of Time Series Data

anomaly-detectiontime-serieslstmarimaunsupervised-learning

Abstraction: Survey of statistical and ML techniques for time series anomaly detection

Key points:

Statistical methods: Holt-Winters (triple exponential smoothing for seasonal data), ARIMA (AR + integration + moving average), HBOS (histogram-based, O(n), good for global outliers but poor on local)
ML supervised methods: Decision Trees (suffer from class imbalance), SVM (robust to outliers via support vectors), LSTM Forecasting (trains on normal data; prediction error = anomaly score)
Unsupervised: K-Means not useful (sensitive to outliers); Hierarchical Clustering preferred with Scipy over scikit-learn for post-hoc cluster selection
Self-supervised: LSTM Autoencoder learns normal-sequence representation; reconstruction error serves as anomaly score
Data representations: point (no context), rolling window (enables contextual detection), time-series features (FFT, CWT, SSA, peak counts)
Key libraries: tsfresh (automatic feature extraction), fbprophet (Facebook's forecasting)

Connections: Anomaly Detection · Time Series · Lstm · Arima