Problems with anomaly detection

Following the hot gym anomaly detection I run it using our own data using a medium swarm but the results do not seem to be as good as when running the data through HTM Studio.

Following interpretation is based on assuming that the anomaly likelihood is correct as specified here: Getting anomaly likelihood

You can see in the following image that even though there is a spike the anomaly score stays pretty much the same. There is a continued high value for likelihood during the spike but closer to the peak than the start (HTM Studio detects the anomaly at the beginning of the spike). Also notice the later prolonged high values for likelihood that do not point at any obvious anomalies in the graph.

In the following image which contains another step there doesn’t seem to be any way to tell that based on either anomaly likelihood or anomaly score. This spike is also detected correctly in HTM Studio.

Checking others I still don’t see any obvious relation between anomaly likelihood/score and an anomaly actually happening.

What can I do so that the adapted hot gym anomaly example will work as well for my data as HTM Studio does?

1 Like