Distance function for ordinal data

[lrebscher]

Hello,

I want to use krippendorff's alpha for ordinal data (6-point Likert scale), but no ordinal distance function is available.

Ordinal distance formula (https://en.wikipedia.org/wiki/Krippendorff%27s_alpha):

Has a support for ordinal data been considered yet?

Cheers

[bryant1410]

@lrebscher I have build a fast version based on this library that contains the ordinal metric. Maybe you can give it a try: https://github.com/pln-fing-udelar/fast-krippendorff

btw, @grrrr thanks for this library!

[bryant1410]

I built it from scratch because this one was too slow for my dataset (it took several hours to compute with 40k units, few annotations per unit)

[grrrr]

Hi Santiago, that's great, will have a look!

Am 28.09.2017 um 05:54 schrieb Santiago Castro notifications@github.com:

I built it from scratch because this one was too slow for my dataset (it took several hours to compute with 40k units, few annotations per unit)

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/grrrr/krippendorff-alpha/issues/5#issuecomment-332720616, or mute the thread https://github.com/notifications/unsubscribe-auth/AAJ-Jp0Xd2G828zb8KIxXtxENzOsurQTks5smxhogaJpZM4Pfirx.

[lrebscher]

@bryant1410 thank you! I will check it out!

[bryant1410]

@grrrr I tested your implementation in https://github.com/pln-fing-udelar/humor/tree/32a7ce954361fb79a58e4e4f282b88a02b4fcda1 To reproduce it:

pip install -r requirements.txt
./agreement.py

Current is in https://github.com/pln-fing-udelar/humor/tree/2ee918293fa99038247f1d848d40bdade471ff0c

I didn't analyze the main bottleneck, but my implementation has some features:

• It computes the coincidence matrix, as the Wikipedia page says to be more computationally efficient.
• It avoids some expensive for loops, trading them by Numpy's tensor multiplications. This is often at least one order of magnitude faster.
• It avoids using try-except blocks inside for-loops, which may be inefficient if it catches an exception very often, as there's a change of context.
• Few helper data structures are used. So it avoids frequent allocations and accesses.
• It uses Numpy wherever possible.

[SaraAmd]

@bryant1410 Is it possible to compute krippendorff's alpha for nominal data with the Euclidean distance function? I am wondering what the distance function is in your fast version.

[bryant1410]

Hey. Now there are distance functions for nominal, ordinal, interval, and ratio data types. Though not sure how you could use Euclidean distance for nominal data.

[SaraAmd]

@bryant1410 Thank you for the response. what are the current distance functions for nominal? how can I see the available distance functions?

[bryant1410]

See https://github.com/pln-fing-udelar/fast-krippendorff/blob/main/krippendorff/krippendorff.py#L14-L39

The distance metric for nominal pretty much checks if the value is the same or different.

[SaraAmd]

@bryant1410 Thank you. Unfortunately, it is not clear what formula has been used. I would recommend having documentation (comments) for a better understanding. Thanks anyway.

[bryant1410]

It follows the definition. See e.g. https://en.wikipedia.org/wiki/Krippendorff%27s_alpha