How to generate multiple ‘extremum’ outliers?

[Bigchicken98]

How to generate multiple ‘extremum’ outliers?

[Bigchicken98]

Thank you for your help

[CodeLionX]

Hey,

You are trying to generate an extremum anomaly of length 15. This is not allowed because extreme values just consist of a single point. If you want to generate multiple extremum anomalies, just add multiple anomaly definitions:

timeseries:
- name: demo
  length: 1000
  base-oscillations:
  - kind: sine
    frequency: 4.0
    amplitude: 1.0
    variance: 0.05
  anomalies:
  # 1. extremum anomaly
  - position: beginning
    length: 1
    kinds:
    - kind: extremum
      min: true
      local: false
  # 2. extremum anomaly
  - &extremum-anomaly
    position: middle
    length: 1
    kinds:
    - kind: extremum
      min: false
      local: true
      context_window: 50
  # 3. extremum anomaly
  - <<: *extremum-anomaly
  # ...

[Bigchicken98]

I adopted your method. I wanted to generate 50 extremum anomalies in base oscillation with length of 1000, but only 32 were generated

[CodeLionX]

50 extreme values in a time series of length 1000 is a contamination of 5%. Don't you think that this is too much for an anomaly detection use case?

Nevertheless, there is only so much that the automatic positioning algorithm in GutenTAG can do. With so many anomalies, you should specify their location manually and don't rely on the "beginning", "middle", "end" positions. You can do this with the exact-position-key:

timeseries:
- name: demo
  length: 1000
  base-oscillations:
  - kind: sine
    frequency: 4.0
    amplitude: 1.0
    variance: 0.05
  anomalies:
  - exact-position: 413  # index beginning from 0
    length: 1
    kinds:
    - kind: extremum
      min: true
      local: false
  # ...

---

I want to have a look why the anomalies disappear. Can you send me your configuration file?

[Bigchicken98]

timeseries-declarations.txt
Thank you!

[CodeLionX]

timeseries-declarations.txt Thank you!

There is actually a bug in the positioning algorithm that leads to multiple anomalies getting injected at the same position. We are working on a fix for that.

[CodeLionX]

We fixed the positioning algorithm in 801b0a3. GutenTAG should not miss any further anomalies from now on. You can use the current master to test this.