Pattern based anomalies in the NAB Data

[ananthmeka]

In the labels.json, we observed that all the labels are based on single instances .

example :
"realTraffic/TravelTime_387.csv": [
"2015-07-30 12:29:00",
"2015-08-18 16:26:00",
"2015-09-01 05:34:00"
]

as per this representation all these anomalies in NAB data are only point based anomalies only ? there is no Pattern Based Anomalies in NAB Data ?

For information : A pattern anomaly (also known as a collective anomaly) is a sequence or group of data points that, when viewed together, are considered anomalous, even if each individual point within the sequence is not. The anomaly is in the pattern of the points, not the individual values.]
Representation of the Point-based anomaly is , since this is an instantaneous event, so it is represented by a single timestamp marking the moment it occurred.
Representation of the Pattern-based anomaly is, since this is a durational event, so it is represented by a start timestamp and an end timestamp to define the entire range of the anomalous pattern.

[subutai]

Pattern based anomalies are actually core to NAB. In fact both temporal and spatial anomalies are important to NAB.

Scoring is based on a window around the event. In addition, for continuously streaming data (the focus for NAB) early detection is critical. Think of heart attacks - you don't want to detect the event itself but weird patterns that indicate the event is about to happen soon. As such, the earlier the system detects the anomaly, the higher the score.

Please see this section in our paper for discussion on the window'ed scoring system: https://www.sciencedirect.com/science/article/pii/S0925231217309864#sec0010