Is Anomaly detection depends on time series , lets 2 cases
Case 1 : without time
x1 ,x2 ,x3 ,x4 …
Case 2 :
(t1 ,x1) , (t2 ,x2) , (t3 ,x3) , (t4 ,x4) …
Are anomaly scores and predictions same for the series of data ?
I moved this from #htm-theory to #htm-hackers because this is not really a theory question.
@wip_user I’m not certain what you are asking, but each row of data should contain one time field that applies to the entire row of data. Not one time per field.
The anomaly score depends on the SDR encodings that are fed into the core SP and TM algorithms. These encodings are formed by the fields that are input to them. Each field (or feature) carries qualitative information that shapes the overall encoding and resulting SDR, so the anomaly score would likely be at least somewhat different for those two cases.
I also want to point out that any data fed into NuPIC is treated as time series data, regardless of whether it has an actual timestamp on it. If there’s no timestamp as in Case 1, then each value is seen in succession to the prior values and NuPIC learns the temporal transitions between them. Although the timestamp value is included in the demos its not actually required. I’d say try both cases with plotting and check out the differences.
That is true. You would only want to include timestamp information if you think there are time-based patterns like “day of week” or “hour of day” in the data set.
Thanks sheiser1 and rhyolight