Figure 9

Comparison of STAR and Twitter’s Anomaly Detection Vector (ADV) algorithm used for detecting phenomena in Twitter 1gram time series. The Jaccard similarity coefficient is presented for each 1-gram and the region where events on detected are shaded for the respective algorithm. Blue-shaded windows correspond with STAR windows of shock-like behavior, while red-shaded windows correspond with ADV windows of anomalous behavior (and hence purple windows correspond to overlap between the two). In general, ADV is most effective at detecting brief spikes or strong shock-like signals, whereas STAR is more sensitive to longer-term shocks and shocks that occur in the presence of surrounding noisy or nonstationary dynamic. ADV does not treat strong periodic fluctuations as anomalous by design; though this may or may not be a desirable feature of a similarity search or anomaly detection algorithm, it is certainly not a flaw in ADV but simply another differentiator between ADV and STAR