tegdet: An extensible Python library for anomaly detection using time evolving graphs

Abstract

This paper presents tegdet, a new Python library for anomaly detection in unsupervised approaches. The input of the library is a univariate time series, representing observations of a given phenomenon. Then, tegdet identifies anomalous epochs, i.e., time intervals where the observations differ in a given percentile of a baseline distribution. Epochs are represented by time evolving graphs and the baseline distribution is given by the dissimilarities between a reference graph and the graphs of the epochs. Currently, the library implements 28 dissimilarity metrics, i.e., 28 different anomaly detection techniques, and its extensible design allows to easily introduce new ones. tegdet exposes a complete functionality to carry out the anomaly detection, through a straightforward designed API. Summarizing, to the best of our knowledge, tegdet is the first publicly available library, based on time evolving graphs, for anomaly detection in time series. Our experimentation shows promising results. For example, Clark and Divergence techniques can achieve an accuracy of 100%, while the time to build the model and predict lasts for few hundreds milliseconds.

Keywords

• Unsupervised anomaly detection
• Univariate time-series
• Time evolving graphs
• Dissimilarity metrics