sktime/sktime

[ENH] minimal episode length smoother for segmentations and change points

Open

#6751 aperta il 11 lug 2024

Vedi su GitHub
 (8 commenti) (0 reazioni) (0 assegnatari)Python (1192 fork)batch import
enhancementgood first issueimplementing algorithmsmodule:detection

Metriche repository

Star
 (7162 star)
Metriche merge PR
 (Merge medio 26g 10h) (86 PR mergiate in 30 g)

Descrizione

Suggested estimator for time series annotation - changepoints and segments.

The algorithm is simple and ensures that episodes, i.e., segments, or time differences between change points, are larger than a given minimum.

A simple algorithm proceeds as follows:

  1. for each segment: check whether length is smaller than min_length; if yes, merge with the next segment, and change labels to the majority of the two.
  2. repeat 1 until no too short segments are found

This is just one possible strategy to do this, another one proceeds as follows:

Loop over segments, shortest first. For each segment, if too short, split in the middle and merge with adjacent segments. Update ordering of lengths each time this is done, until no too short segments remain.

The syntax would be that of a compositor: EnsureMinLengthSegments(my_annotator, strategy="strategyname")

Guida contributor