Hi all,

Inspired by Ted Dunning's talk on Anomaly Detection in [1], we are hoping
to implement this in Siddhi.

Basically, this is how Delta Anomaly Detection works.

a. Use a predictor based on percentiles which predicts the next event based
on historical data.
b. Obtain the delta between the predicted value and the actual next value
(delta = Predicted Value - Actual Value)
c. Observe the distribution of delta and flag events where delta is larger
than a threshold percentile.



So basically, for demonstration purposes, the idea is to use the Time
Series Regression functionality and create a predictor for some event
stream. (Any ideas for an interesting event stream?)
Then we compute the delta using siddhi queries. Establish the distribution
of delta, by running a historical dataset through the above process. Then
compare the incoming delta with the percentiles of deltas received so far.
The percentile extension created by Nishali can be used for this. The
incoming delta values will also update the distribution in real time.

Thoughts?

Cheers,
Seshika

1. http://www.slideshare.net/tdunning/strata-2014-anomaly-detection
_______________________________________________
Architecture mailing list
[email protected]
https://mail.wso2.org/cgi-bin/mailman/listinfo/architecture

Reply via email to