Selection window size for Anomaly Likelihood

Hi , i want to use Anomaly Likelihood in order to test a data stream with length of 3000 seconds.I read in papers that W << W ’ and also some values for windows, that refer to longer data sets.What would be some good sizes, for W and W’ in order to test a data stream with this length ?

I believe the default window is 500 inputs. But this window does not affect how long memory is persisted, it only affects how far back the anomaly likelihood calculation goes. I’m not sure changing it will do what you think. (CC @subutai)

1 Like

Thanks for the answer. The W’ has to be really small ? If for window W size is 500 inputs, for W’ should it be 5 inputs?

Sorry I don’t know the answer to your question. Maybe @scott or @subutai does.

I believe the default values are 10 for the averaging window and very large (>8000) for the historical window to compare against. I strongly recommend using these defaults first. You could try different averaging windows (from say 5 to 30) but the historical window size isn’t likely to have much of an impact (you could try going down to as little as 500 or 1000).

I think these parameters are probably not going to make or break the performance of your system. The encoding and frequency of measurements will have a much bigger impact. You could try a small amount of aggregation if your data changes slowly. But if your data is changing slowly then 3000 records probably isn’t enough for the algorithm to learn all of the different types of sequences that can occur.

2 Likes