Graph-Based Time Series Anomaly Detection
Lepley, Amanda · Digital Repository at the University of Maryland (University of Maryland College Park) · 2015
Time series anomaly detection is an important problem in the field of machine learning. To solve this problem, we propose a method for detecting anomalies in time series data using a graph-based algorithm. The algorithm first constructs a graph model of the various states in the data by clustering groups of points as they come in, and labeling the edges between the states in the graph as anomalous or normal. After this training phase, as new data comes in the algorithm determines whether or not segments of data are anomalous based on features of the edge that connects the new data to the graph. This algorithm has been shown to work on real data from the medical realm, as well as synthetic test data. The algorithm has the advantage of being very efficient. It can work faster than real time in a streaming fashion for data sampled at 256 Hz.