add data drift example notebook for text data

[SangamSwadiK]

Related to :
#392

What does this implement?
This adds an example notebook for detecting data drift for text data using evidently

• add data drift for text data
• add relevant headings
• explanation at each step
• references and citation for datasets used

Summary of things in the example notebook :

1. I've used averaged glove vectors representing each sentence(record) across, IMDB movie review train , test and amazon echo-dot review datasets. I've used Glove 50dim for calculating the average vector for each record

2. I've created visualization based on Dashboard and created drift measurement based on TestSuite using wasserstein distance for numerical features.

3. Dataset used for reference (train and test): IMDB movie reviews.
   Dataset used for current(Unseen dataset for drift comparison) : Amazon Echo-Dot reviews

Please let me know if improvements could be made in measuring data drift for text. Not completely sure if this is the right approach.
I tried using the word counts after and using Chi-squared test on the word counts between train and test, and
train and unseen datasets, It didn't seem to work well.

I'm reading the data from my raw.githubusercontent, please let me know if there is a better way to store and read data from, or perhaps data could be stored in a separate folder.

Attaching a link to colab notebook.
https://colab.research.google.com/drive/1Y5X-MkZELIQv3-OHeWZTyyfftE5Lju58?usp=sharing

Thanks !

[emeli-dral]

Hi @elenasamuylova,

I believe you would love to have this example 😎

[emeli-dral]

@SangamSwadiK we are exited to have an example with text data! Thank you.

Are you going to add more commits to this PR? I marked it as a hacktoberfest-accepted, to give you a chance to add something if you want.
And I kindly ask you to move your example from "sample_notebooks" to "how_to_questions", cause it fits there perfectly. It answers a question "How to calculate drift on top of the text data".

[SangamSwadiK]

@SangamSwadiK we are exited to have an example with text data! Thank you.

Are you going to add more commits to this PR? I marked it as a hacktoberfest-accepted, to give you a chance to add something if you want. And I kindly ask you to move your example from "sample_notebooks" to "how_to_questions", cause it fits there perfectly. It answers a question "How to calculate drift on top of the text data".

Hi ! @emeli-dral
Yes, I think there would be more commits, primarily, because of two questions ;

1. How would you suggest to handle loading glove vectors?
   • I've commented the downloading and unpacking of glove vectors because
     • I thought it might take more time to run the github workflow
     • And it could be easily done locally, and the user need not be using glove vectors. They could be using something else. That's the reason I commented it.

2. The data used (Imdb reviews and amazon eco-dot reviews) are from raw.githubusercontent from https://github.com/SangamSwadiK/test_dataset
   I think it would be better if this data were in Evidently repository maybe under a directory evidently/examples/how_to_questions/data(or any other better name). Please let me know if there is any other way to handle this.

Other things to be changed ;
I think the notebook could be cleaned further, add more detailed comments at each step and other minor changes.
I agree, moving it to how_to_questions because it answers a question.

what do you suggest ?
Thanks.

[SangamSwadiK]

Updated the notebook (added relevant comments and credits) and added the sample data as well.

[emeli-dral]

Hi @SangamSwadiK!

I take a loot to the dataset.
1. loading glove vectors
I agree with your approach. If a user decides to run it locally, it can easily uncomment the downloading and unpacking code and run it.

2. dataset in evidently/examples/how_to_questions/data
We discussed that, and we decided against storing data in the repo directly. The main reason is that we do not want to have our data collection now and support it further (especially when it comes to storing many datasets which can originally be distributed under variety of licences).
I suggest to add a link to the original dataset together with the comment on how this data were processed to get https://github.com/SangamSwadiK/test_dataset, and keep using test_dataset. What do you think?

[SangamSwadiK]

Hey @emeli-dral !
I've updated the notebook. I've commented out the glove vector download. I've also added an introduction, instruction section along with necessary links to data, preprocessing and my preprocessed data(github repo).

Please let me know in case of improvements. Will fix them asap.
Thanks for reviewing!