Hero research featured on the cover of Entropy
In the article by Sabeti et al., a new information theoretic anomaly detector for time series is introduced. The method is based on detecting changes in the compressability of a test segment of the time series as measured by the difference between complexities of a typical encoder and a universal encoder. The typical and universal encoders are respectively implemented with a tree-structured pattern dictionary, trained on an earlier segment of the time series, and a Lempel–Ziv encoder. The anomaly detector is illustrated for a chaotic time series with model shift and for early detection of anomalous heart rates and skin temperatures of patients after exposure to a respiratory virus.
Read the article, A Pattern Dictionary Method for Anomaly Detection, by Elyas Sabeti, Sehong Oh, Peter X. K. Song, and Alfred O. Hero.