Tutorial: Deep Learning Methods for Unsupervised Time Series Anomaly Detection

Abstract

Time series anomaly detection (TSAD) is a crucial tool for detecting unusual patterns or outliers in large-scale time series datasets. With the increasing amount of data being generated in various industries, from financial transactions to industrial processes, it has become essential to identify anomalies that may indicate unusual events, fraud, or errors in the system. Deep learning algorithms have proven to be effective in detecting these anomalies, as they can learn complex patterns and relationships in the data that may not be apparent to humans. By detecting anomalies early, businesses can take proactive measures to prevent or minimize their impact, improving their overall performance and reducing risk. Thus, deep TSAD is an important tool for ensuring large-scale systems’ quality, reliability, and security and has attracted huge interest from the research community. This tutorial aims to provide the audience with a comprehensive and organized tutorial on state-of-the-art algorithms for time-series anomaly detection. We will first discuss the basic concepts of anomaly detection and time-series processing. Then, we will thoroughly review the deep learning methods for time-series anomaly detection, ranging from traditional autoencoder-based algorithms to the recently-proposed graph-based models. In the final parts of this tutorial, we will focus on some important considerations during the implementation, evaluation and dataset preparation for the deep time-series anomaly detection algorithms. Besides introducing the SOTA methods, this tutorial will aim to enable the researchers to apply them to solve real-world problems involving time series.

Narges Armanfard (Chair)

Assistant Professor

Narges Armanfard is a Tenure-Track Assistant Professor of Electrical and Computer Engineering at McGill University, and Mila - Quebec AI Institute. She is also affiliated with McGill Centre for Intelligent Machines (CIM), McGill initiative in Computational Medicine (MiCM) and McGill Institute for Aerospace Engineering (MIAE). Before joining McGill, she was a postdoctoral researcher in the Intelligent Assistive Technology and Systems Lab at the University of Toronto and University Health Network. She obtained her PhD degree in Electrical and Computer Engineering from McMaster University.

Hadi Hojjati

PhD Student

Hadi received his Bachelor’s degree in Electrical Engineering from Sharif University of Technology, Iran. He started as an MSc student at the iSMART Lab. and fast-tracked to PhD. He has received Graduate Excellence Fellowship Award (GEF), McGill Engineering Doctoral Award (MEDA), GREAT Award, and AGE-WELL Award. He is currently doing research on multi-modal and multi-variate anomaly detection in time-series data.

Thi Kieu Khanh Ho

PhD Student

Khanh received her Master’s degree at Gwangju Institute of Science and Technology (GIST), South Korea. She was a researcher at Seoul National University Hospital (SNUH), South Korea, before moving to Montreal for her PhD. She has received the prestigious McGill Engineering Doctoral Award (MEDA), GREAT Award and is nominated for the prestigious Vanier Scholarship. Her research focuses on time-series data anomaly detection using generative models and graphs.