/Anomaly_Detection_TadGAN

Houses code for an anomaly detection pipeline that can detect anomalies in time series data in real-time using TadGAN, Spark and Kafka

Primary LanguageJupyter Notebook

Anomaly_Detection_TadGAN

Houses code for an anomaly detection pipeline that can detect anomalies in time series data in real-time using TadGAN, PySpark and Apache Kafka.

Introduction

Time series anomalies can offer information rel- evant to critical situations facing various fields, from finance and aerospace to the IT, security, and medical domains. We apply one such model, TadGAN (Geiger et al., 2020) which is built on top of GANs and is meant for unsupervised anomaly detection in time-series data. TadGAN has already improved performances on six out of eleven benchmark datasets such as NASA, Yahoo- S5 and NAB. All of the benchmark datasets of TADGAN are univariate i.e a single signal is being used to re- generate single signal. We extend its use to other datasets across domains using the multivariate SKAB dataset and compare results obtained by TADGAN with other anomaly detection models.

Tech Stack

  • Python 3
  • PySpark
  • Apache Kafka
  • Scikit-Learn
  • Orion-ML
  • PyOD
  • Pandas

Anomaly Detection Pipeline

Report and Presentation

You can access the report here and presentation here