TffvVectorizer Enconding

[DelgadoPanadero]

With this PR I have implemented the TffvVectorizer class. This class is a text feature encoder that works similarly to TfidfVectorizer, however it performs better results in the problem of imbalanced class categorizacion, as it can be seen in this study

Liu, Y. et al., Imbalanced text classification:
doi:10.1016/j.eswa.2007.10.042

Personally I have checked the improvement in the performance of this encoding compared with the classical tfidf encoding for tweets classification. It would be a pleassure for me to contribute to sklearn, so any feedback would be great. Thank you!

[jnothman]

I don't think this qualifies for inclusion in scikit-learn (see our FAQ for rough criteria). Consider publishing it elsewhere, such as through Scikit-learn-contrib

[rth]

Yes, please make this PR to https://github.com/scikit-learn-contrib/scikit-learn-extra/ . Also I think the transformer would be enough, no need to add the vectorizer (see #14966).

[rth]

To elaborate why this doesn't qualify for inclusion at the moment: there are a number of unsupervised or supervised weighting schemes that are proposed on top of TF-IDF (see e.g. https://github.com/textvec/textvec and also TF-IGM) for term weighting in text classification. Most claim to over-perform TF-IDF in some cases. Whether there are some that are consistently better / are generally accepted as such by the community at the moment is not clear, and that's why it would be better to start with adding this PR to scikit-learn-extra, then considering adding it to scikit-learn in the future if the user feedback is very positive.

[DelgadoPanadero]

Ok I see your point, thank you for your feedback and congratulation for the quickness in the answer, I didn't expect it to be so soon. I have a doubt. When tagging the issue with 'Move to scikit-learn-extra', do you mean that I should add the PR to scikit-learn-extra or you task to add it when you consider?

I wil think about other enhancements for the library, because I would really like to help to improve it. Thank you!

[rth]

I made a related PR recently scikit-learn-contrib/scikit-learn-extra#45 . I would be good to use the same benchmarks to compare once that one is merged (and at least you should put you estimator in the same module).