Abstract: An incremental-decremental algorithm adds and removes points from a sample/training set in an efficient way - typically with finite memory overhead, and finite computation for each add/remove operation. We will explain some incremental-decremental methods we developed for the purpose of maintaining rolling window statistics. The techniques include a novel representation of a computation as a DAG, and adaptations of ideas from econometrics. As a consequence, we expand the range of anomaly detection, forecasting, and time series classification that can be performed in the streaming setting.
Our methods were motivated by use cases from infrastructure monitoring, and certain UX and platform requirements for a data-intensive monitoring product. The talk will provide some business context.
Bio: Joe Ross holds a PhD in mathematics from Columbia University and was a researcher and instructor in pure mathematics, most recently at the University of Southern California. He has worked as a data scientist at machine learning/analytics startups for several years; in his current role at SignalFx, he focuses on a variety of time series problems.
