Anomaly Detection in Streams with Extreme Value Theory

By Alban Siffer, Pierre-Alain Fouque, Alexandre Termier, Christine Largouët. Knowledge Discovery and Data Mining, 2017

606 citations recorded by the source corpus.

Subfields: Artificial Intelligence, Computer Networks and Communications

View Paper: https://www.semanticscholar.org/paper/0cfe631a0f99242a751a5a8bfd28dfb18f37f23e

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Computer Networks and Communications

Citations in the corpus

A Review on Outlier/Anomaly Detection in Time Series Data (2020, methodology): “Additionally, they apply extreme value theory [Siffer et al. 2017] in the reference of normality to automatically select the threshold value.”
Time-Series Anomaly Detection Service at Microsoft (2019, methodology): “For example, Holt winters [5] always shows poor results in (b) and (c); and Spot [19] always shows poor results in (a).”
TS2Vec: Towards Universal Representation of Time Series (2021, methodology): “L 1 distance is used to measure the anomaly score: To avoid drifting, following previous works (Ren et al. 2019), we take the local average of the preceding Z points We compare TS2Vec with other unsupervised meth-ods of univariate time series anomaly detection, including FFT (Rasheed et al. 2009), S”
An Evaluation of Anomaly Detection and Diagnosis in Multivariate Time Series (2021, background): “Time-series anomaly detection techniques may be categorized based on their detection technique, namely, shallow model-based [19]–[22], deep-learning model based [3]–[6], [8]–[11], [14], [17], [23], pattern-based [24], [25], distance based [26], [27] and non-parametric [28].”
Multivariate Time-series Anomaly Detection via Graph Attention Network (2020, methodology): “We use Peak Over Threshold (POT) [28] to choose the threshold automatically.”
Learning Graph Structures With Transformer for Multivariate Time-Series Anomaly Detection in IoT (2021, methodology): “OmniAnomaly [13] was a stochastic model to avoid potential misguiding by uncertain instances, which used stochastic variable connection and normalizing flow to get reconstruction probabilities and adopted streaming POT with drift (DSPOT) algorithm [38] for automatic”
Graph Neural Network-Based Anomaly Detection in Multivariate Time Series (2021, methodology): “While different approaches could be employed to set the threshold such as extreme value theory (Siffer et al. 2017), to avoid introducing additional hyperparameters, we use in our experiments a simple approach of setting the threshold as the max of A s ( t ) over the validation data.”
Deep Learning for Time Series Anomaly Detection: A Survey (2022, neutral): “Common methods for establishing these thresholds include Nonparametric Dynamic Thresholding (NDT) [92] and Peaks-Over-Threshold (POT) [158], which help convert scores into final labels.”
Anomaly Detection in Time Series: A Comprehensive Evaluation (2022, methodology): “[91] SubsequenceLOF [22] Data Mining TwoFinger[90] NorM[14] DissimilarityAlgo[6] STOMP [164] Series2Graph [16] HOTSAX [70] STAMP [156] MERLIN[97] DAD[154] BoehmerGraph[13] GrammarViz [120] TARZAN [71] NormA-SJ [15] NormA-smpl[15] KnorrSeq2[102] TSBitmap [144] ILOF[108] VALMOD [82] SurpriseEncoding[2”
Practical Approach to Asynchronous Multivariate Time Series Anomaly Detection and Localization (2021, methodology): “Some authors were aware of this problem and suggested approaches to approximate threshold selection through dynamic error smoothing over a time window sequence [13], or through applying the principle of Extreme Value Theory (EVT) [26] on the fitted probability distribution for the Bayesian model [28”
TranAD: Deep Transformer Networks for Anomaly Detection in Multivariate Time Series Data (2022, methodology): “Anomaly diagnosis label for each dimension 𝑖 (𝑦𝑖 ) and detection (𝑦) results is defined as 𝑦𝑖 = 1(𝑠𝑖 ≥ POT(𝑠𝑖 )), 𝑦 = ∨ 𝑖 𝑦𝑖 .”
Robust Anomaly Detection for Multivariate Time Series through Stochastic Recurrent Neural Network (2019, methodology): “Peaks-Over-Threshold (POT) [22] is the second theorem in EVT.”