Univariate Time Series Anomaly Detection Based on Variational AutoEncoder

EasyChair Preprint 8445

Abstract

In the field of anomaly detection, the boundaries of anomalies are always blurred, and professional knowledge is required to define them, which consumes a lot of manpower and time to mark what anomalies are. In this paper, a Variational Auto-Encoder(VAE) neural network model is used, and an unsupervised learning anomaly detection model that considers both temporal dependencies and reconstructed features. In the calculus of marking outliers, we propose a two-dimensional sliding window with a clustering algorithm to solve the traditional method of judging outliers using a single threshold. Experimental results based on Yahoo Webscope dataset show that the performance can be ameliorated by the proposed method.

Keyphrases: anomaly detection, two-dimensional sliding window, variational auto-encoder

@booklet{EasyChair:8445,
  author    = {Leehter Yao and Youwei Chang},
  title     = {Univariate Time Series Anomaly Detection Based on Variational AutoEncoder},
  howpublished = {EasyChair Preprint 8445},
  year      = {EasyChair, 2022}}