DeepAD: A Generic Framework Based on Deep Learning for Time Series Anomaly Detection

Abstract

This paper presents a generic anomaly detection approach for time-series data. Existing anomaly detection approaches have several drawbacks such as a large number of false positives, parameters tuning difficulties, the need for a labeled dataset for training, use-case restrictions, or difficulty of use. We propose DeepAD, an anomaly detection framework that leverages a plethora of time-series forecasting models in order to detect anomalies more accurately, irrespective of the underlying complex patterns to be learnt. Our solution does not rely on the labels of the anomalous class for training the model, nor for optimizing the threshold based on highest detection given the labels in the training data. We compare our framework against EGADS framework on real and synthetic data with varying time-series characteristics. Results show significant improvements on average of 25% and up to $40-50$% in $F_1{\text{- }}score$, precision, and recall on the Yahoo Webscope Benchmark.

This project has received funding from the European Union’s Horizon 2020 research and innovation programme under the Grant Agreement No. 700381 (ASGARD) and No. 671625 (CogNet).