[Question] Can I use a List of sequence indices as an Input?

[Ishitori]

I have a feature column which is an array containing list of items a user has interacted with. I found a Kaggle example with a similar situation where the author of the notebook uses this kind of feature: https://www.kaggle.com/code/matanivanov/wide-deep-learning-for-recsys-with-pytorch

Is it possible to do the same thing using Pytorch-widedeep library?

[jrzaurin]

I will have a look to the Kaggle example and write back, but as long as you can replicate the output of the TabPreprocessor (which is fairly standard) you can pass anything that has that format.
I will try to write an example later with a column in that is a list of tokens (which I think is what you need)

[Ishitori]

Yes, you can think of this as a list of tokens. It is usually something as a history of purchases, or as in Kaggle example, a list of already watched movies. The order might matter or might not. In my case it does (items are ordered by decreasing of frequency of interaction).

[jrzaurin]

@Ishitori sorry I did not have time today, I think it will have to wait until Fday. But if order matters, then a "standard" WideDeep architecture where the deep component is an MLP wont work, right? in that case you would prefer an RNN or a transformer? (note that at the moment we do not support Transformer-based models for the text component. My next updgrade to the library is integrate it with Huggingace.

[Ishitori]

Oh, sorry for the delayed reply. I would prefer Transformer (attention-based) model, and it doesn't need to work with text (I don't have text features).

Thank you!

[ncsilvaa]

Hey everyone! This is an interesting feature for recommender systems.
Could you provide an example of how to use it?
Is something possible to do with the current version of the library?
Thanks

[jrzaurin]

Ill comment on this asap, travelling at the moment 🙂

[AresDan]

Any progress on this topic? Faced similar issue

[jrzaurin]

I have holidays approaching soon and I will address some of the open issues, including this :)

thanks!

[jrzaurin]

hi @Ishitori @AresDan and @ncsilvaa

I will soon open a PR form the wide_deep_recsys branch. In the meantime, you have an example of how to use the library for recsys following the notebook that @Ishitori mentioned in his first comment. All the code is here: https://github.com/jrzaurin/pytorch-widedeep/tree/wide_deep_recsys/examples/scripts/wide_deep_for_recsys

I still have to add some unit tests and I will also include a notebook with more explanations here and there. But in the meantime you can have a look to that code which I hope answers the questions here.

Any other question, please, let me know :)

[jrzaurin]

PR #183 addresses this issue