Some New Thoughts on Anomaly Detection and Time Series Modeling

Anomaly Detection methods are a much-discussed topic. Every application wants to detect unusual behavior, for reasons from reliability to security to performance. The reality is that with so much data, such detection methods need to be able to be processed in real-time, and most applications don’t have time for an algorithm to “learn” the system. Moreover, traditional anomaly detection algorithms rely too much on initial bias and human opinion, while building on a relative definition of what an anomaly is. In this talk, we’ll discuss the issues with conventional anomaly detection methods, and introduce some ongoing research into the simplification of time series modeling by transforming the mathematical space in which the data resides, allowing aggregation of the data that eschews point estimates in favor of retaining more information. We discuss how this will build better anomaly detection methods, and list some applications of such methods in storage.

Learning Objectives:
1. Learn what a fuzzy number is
2. Learn about formal definition of time series anomalies, or interventions
3. Learn the shortcomings of current anomaly detection algorithms
4. Learn about fuzzy ARIMA modeling and its implications for better anomaly detection in storage applications