[Joss Review] Comparison to tslearn/seglearn about seg1d

[AKuederle]

Sry I just realised you already mentioned tslearn in the README. It would still be interesting to have some further information under which circumstances this package should be chosen over the existing alternatives

[cadop]

I added an additional explanation in the README for this topic, but I also will try to address some of it here as well.

First, I do use some methods in sklearn, so there is overlap there, but it is in order to achieve the goal of seg1d.

I think there is also some nuance to the packages and wording that conflates some of the differences. I spent a few hours trying to get seglearn and tslearn running so that I could compare them, but reached a point that I either do not understand what they are doing enough to modify their examples, or the differences are user-related. For example, the tslearn method you linked to is using cross correlation (seg1d uses rolling correlation), and is clustering the time series (seg1d uses clustering, but it is really just to determine which segments are the best match, not for classification).

The closest method I can find in tslearn is this example: https://tslearn.readthedocs.io/en/stable/auto_examples/misc/plot_distance_and_matrix_profile.html#sphx-glr-auto-examples-misc-plot-distance-and-matrix-profile-py

Essentially the distance profile is similar to the output of the rolling window correlation. Other than that, it is using the minimum distance to find the segment whereas in seg1d, it uses peak detection. The clustering part would be for example, the neighbor dist that is shows with a red x in their figure by finding all the x's and clustering which one is the desired motion. The other difference is how seg1d can use both multiple features and multiple sets, along with weighting values. In the end, I can see how to get similar results, although it requires a good deal of effort (i.e., how seg1d ended up being created). Other small features of seg1d I don't think are too important in this matter.

I also realize I can be missing some essential info, so if you know of a way that these libraries are facilitating the same process/outputs as the API examples please let me know, I can also test the differences in output/usability as well then.
These would be the two cases:

https://cadop.github.io/seg1d/build/html/api_ecg.html

https://cadop.github.io/seg1d/build/html/api_feat.html

I used the ecg example data from seg1d on the tslearn and got to this point

It doesn't have the multifeatures/references but shows a similar distance profile. However, this is as close as I could get to seg1d based on their examples.

[AKuederle]

Thank you very much for putting so much effort into this! I guess you are right, non of packages methods includes a rolling correlation based system for segmentation. There are other methods (e.g subsequence DTW in tslearn), but I agree that there are not identical the same.

From a scientific point of view, could you quickly explain what die advantages of rolling correlation + peak detection compared to normalized cross correlation + peak detection is? I would consider the latter approach pretty popular.

[cadop]

The result of a rolling pearson correlation is the same as a normalized cross correlation implementation. NCC is not the same as cross correlation, and based on what I expect most users to be confused with, its easier to refer to rolling correlation since the common methods of np.correlate(x, y, 'full') is not NCC, but cross correlation.