Anomaly Likelihood does not work correctly!

[Thanh-Binh]

Hi all,
I have just tested anomaly detection with sinus waves by using

1. TM with embedded Anomaly Likelihood
2. TM with embedded Anomaly disable + external AnomalyLikelihood class
   For both cases, I used default values of AnomalyLikelihood.

In my observation:

1. one-step-prediction provides very good prediction data, (good)
2. raw anomaly has more and more small anomaly scores ...(good)
3. BUT anomaly likelihood jumps between 0.5 (for some time period) to 1 (for other time period) and back to 0.5 and so on.

Does anyone have something like my observation? Thanks

[Thanh-Binh]

By comparing my codes ported originally from python version, this version is totally different! This explains me!
I see that TM.anomaly seems to be likelihood, the anomaly score should be
1.0 - TM.anomaly

[breznak]

Related to review of anomaly likelihood #469

I have just tested anomaly detection with sinus waves by using

@Thanh-Binh could you share your simple experimentation code for sine waves? I'd like to integrate it to our tests as quality benchmark for different anomaly modes.

[Thanh-Binh]

@breznak I will share it asap...

[breznak]

This could likely be joined with #646 , as the experiment on sines is the same, and we'd do both predictions & anomaly.

AFAIK the hotgym runs sine wave, @Thanh-Binh if you have problems porting your codes, can you explain on the existing hotgym where the problems are?

[breznak]

I see that TM.anomaly seems to be likelihood, the anomaly score should be
1.0 - TM.anomaly

I was able to confirm this, escalating to a bug. Thank you for reporting!

I was confused by "likelihood of anomaly" (which would imply high likelihood => anomaly), but it is a likelihood of a score (anomaly score). And we want to detect unlikely scores -> 1.0 - likelihood is correct anomaly interpretation here.

[breznak]

See #469 (comment) for some high level ideas about anomaly likelihood.