Detect patterns/outliers using WarpScript

[Pierre Zemb]

Hi everyone!

As I am digging more and more into WarpScript, I'm wondering what can I do to detect pieces of data inside a GTS. By reading the documentation, I highlighted theses functions:

• PATTERNS function
  
• THRESHOLDTEST function and the others outliers functions
  
• DTW function
  

As far as I can see, a spike is easily found with outliers functions. I tried to use PATTERNS to detect a specific piece of GTS, with mitigated results. I guess it depends of what I'm looking for, but is there some best practices to detect:

• A spike?
• A piece of GTS which is "flat", or the opposite, (i.e with high fluctuation)
• A very specific piece of GTS?

Are there any other functions ?

PATTERNS documentation is talking about these parameters:

• pattern length must divide window size

• quantization scale must be a power of two from 2 to 65536

• PATTERNS can only be applied to numeric Geo Time Series which are bucketized and filled (i.e. no bucket with missing values).

What's the unit of pattern length? In points, microsecond? Do you have guidelines to set the PATTERNS parameters?

[Mathias Herberts]

The window width for PATTERNS is in time units.

The best way to detect a 'flat', i.e. constant, part of a GTS is probably to compute the delta and use a threshold.

For detecting specific shapes, either PATTERNS or DTW can be used, the actual results will depend on noise, time warping and similarity among subsequences of the GTS.

[Pierre Zemb]

Thanks Mathias for the tips!