Graph-based anomaly detection: When and how

Description

Thanks to frameworks such as Spark’s GraphX and GraphFrames, graph-based techniques are increasingly applicable to anomaly, outlier, and event detection in time series. However, most data do not naturally come in the form of a network that can be represented in graphs. Therefore, it is not clear whether graph-based techniques always offer the most appropriate approach to detect anomalies.

Jeffrey Yau offers an overview of applying graph-based techniques and outlines the benefits of graphs relative to other techniques. Jeffrey compares and contrasts the use of graph theory and techniques, large-scale time series mining methods, and traditional parametric linear and nonlinear time series techniques in anomaly, outlier, and event detection—with specific examples from credit card fraud, wearable IoT devices, and financial time series.

Topics include:

• Static graphs
• Dynamic graphs
• The most common large-scale time series mining methods
• Traditional parametric linear and nonlinear time series techniques, including change-point detection
• Trade-offs need to be made when applying each of these classes of techniques to identify anomalies