I'm not a statistician, which is why I really like this kind of approach. It feels "FiveThirtyEight" trustworthy to me – a weighted average of different opinions.
The claim that it "dynamically adjusts the weight of each algorithm according to the amount of information it contains" is a bit hyperbolic, though. The "weight" function is explained in the code this way:
"If either the "magnitude" [or] "fence" methods don't have any probability to contribute, we don't want to hear about it. If they do, we upweight them substantially."
Regardless, this is unquestionably a great foundation to build into your own systems. Perhaps not as turn-key as the blog suggests, but still a wonderful contribution.
Has anyone played with Anomalyzer yet? I'd like to get the anomaly probability for every data instance of a CSV but it's unclear to me if this is possible. I want to run it on the Numenta Anomaly Benchmark dataset (https://github.com/numenta/NAB) to see how it compares to other anomaly detection algorithms.
7 comments
[ 2.3 ms ] story [ 19.4 ms ] threadThere's a fair amount of user input here (still) - picking weights and thresholds is a nontrivial task.
If you're interested in the changepoint detection use cases, we have a package for that, too. :)
https://github.com/lytics/impact
The claim that it "dynamically adjusts the weight of each algorithm according to the amount of information it contains" is a bit hyperbolic, though. The "weight" function is explained in the code this way:
"If either the "magnitude" [or] "fence" methods don't have any probability to contribute, we don't want to hear about it. If they do, we upweight them substantially."
Regardless, this is unquestionably a great foundation to build into your own systems. Perhaps not as turn-key as the blog suggests, but still a wonderful contribution.