[FEATURE] Add embeding ranker in PyTorch

[miguelgfierro]

Description

https://www.tensorflow.org/recommenders/examples/basic_ranking, but using PyTorch instead of TF

Expected behavior with the suggested feature

Willingness to contribute

• Yes, I can contribute for this issue independently.
• Yes, I can contribute for this issue with guidance from Recommenders community.
• No, I cannot contribute at this time.

Other Comments

[jmarrietar]

Hi @miguelgfierro 👋 , do you mean implementing class RankingModel as in the recommenders example? Could you elaborate a little bit ? :)

[miguelgfierro]

it would be to have a similar model, but in PyTorch. There needs to be code in recommenders folder and a notebook showing an example. If someone wants to implement this, please let me know first because the example of TF is not constructed in the way we do at Recommenders.

[jmarrietar]

Hi Miguel, I'm not quite familiar yet with recommenders structure, but would like to give it a try and implement this😃

[miguelgfierro]

Nice @jmarrietar

Feel free to attend the next reco meeting on Monday if you want the core team to give you some ideas.

Here some details on how to implement this:

1. Try to use a similar structure as in most notebooks: https://github.com/recommenders-team/recommenders/blob/main/examples/00_quick_start/sar_movielens.ipynb load data, split, instantiate model, train model, predict model and evaluation.
2. For loading movielens you can use movielens.load_pandas_df
3. In the TF page there are two classes: RankingModel and MovielensModel, this is a great example of how not to code. There should be a model called NNEmbedding or whatever you want that defines the structure in the __init__ and then you have a fit method that would input the data and training parameters model.fit(train_df, batch_size, epochs...)
4. There is no need to do the serving part.
5. The evaluation can be done similarly to the SAR notebook.

Let me know if you have any questions.

[srivani21]

I can work on this enhancement, I am thinking that this is based on User and Item like Matrix Factorization and Softmax example, so do I need to implement this based on MF and Softmax or any other but by considering user_id, movie_id and rating?

[miguelgfierro]

@srivani21 this issue is not for matrix factorization, it's a fully connected network.
If you are interested in MF, there are two opportunities to collaborate: #1713 and #2186

[jmarrietar]

Thanks @miguelgfierro , I will try to attend the Monday meeting to gather ideas, Thanks!