feat(#1134): Allow extending the weak label matrix with embeddings

[dcfidalgo]

Closes #1134 and #1309

[dcfidalgo]

Ok, I will merge this tour de France when the tests pass.

A quick conclusion of this journey:

• We figured out why Ruan had problems in the beginning with the ag news dataset. Since the dataset is big, he computed the embeddings in colab, but in the order of the original dataset from the HF Hub. In the tutorial we upload the records without an id, so the order in which the weak label object retrieves the records is random. So there was a mismatch between the records and the embeddings, that's why he basically obtained a random classifier. I will open an issue with a proposal to mitigate this issue in the future.
• Optimizing the threshold is important, so it is not necessarily a cheap approach to improving performance. We could do some more experimentation to find reliable and cheap optimization methods.
• In general, when one wants to extend the weak label matrix, one should aim for rules with very high precision and not necessarily a high recall, since this is what the extension of the weak label matrix is for. So one has to keep this in mind right from the start.
• Also, the right embeddings are obviously important as well.

CC @dvsrepo @frascuchon

[frascuchon]

Well, it looks like we're finally getting this feature...

As improvement points:

1. The embeddings should be part of the records in the weak label matrix, instead of being passed as an argument in the extension. This can remove potential mismatch embedding errors.

for record in wl.records():
   record.embedding = .....

So, the extend_matrix method will accept only the thresholds argument.

2. It would be nice to provide some default method or helper to optimize the thresholds that are stopped for the extension, given the importance of this parameter.

But it should be better to tackle them (if apply) in a different PR

[codecov]

Codecov Report

Merging #1487 (b23f95c) into master (e8f5613) will increase coverage by 0.00%.
The diff coverage is 96.42%.

@@           Coverage Diff           @@
##           master    #1487   +/-   ##
=======================================
  Coverage   95.52%   95.52%           
=======================================
  Files         126      126           
  Lines        6232     6278   +46     
=======================================
+ Hits         5953     5997   +44     
- Misses        279      281    +2     

Flag	Coverage Δ
pytest	95.52% <96.42%> (+<0.01%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

Impacted Files	Coverage Δ
...rubrix/labeling/text_classification/weak_labels.py	99.34% <96.42%> (-0.66%)	⬇️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update f6ff8dc...b23f95c. Read the comment docs.

[dcfidalgo]

I just talked about his comments with @frascuchon

The embeddings should be part of the records in the weak label matrix, instead of being passed as an argument in the extension. This can remove potential mismatch embedding errors.

We think it would be very helpful if one could materialize/load a WeakLabels object. Then you can easily share your dataset + weak labels, load it on a machine with GPU to add embeddings, etc.

It would be nice to provide some default method or helper to optimize the thresholds that are stopped for the extension, given the importance of this parameter.

I think more tests with different datasets and rules would be necessary to find a reliable and practical (cheap) method ... would be a good project for a student :)