Item-to-item recommendation

[max-grzanna]

The quick start guide says TFRS can also be used for item-to-item recommendations. In the retrieval guide, it says: "... and train the model using (query item, candidate item) pairs." If, for instance, my training data looks like this (containing product IDs from bought together items):

product_one	product_two
22	4442
22	7763
125986	1252
...	...

should the candidate tower only contain all the unique items from column two, or all the unique items across the whole dataset?

How would a dataset look like for an item-to-item recommender?

Any response will be highly appreciated! :)

[maciejkula]

That's up to you!

One way to fit a model like this is to have a shared embedding table (and vocabularies) between both the query and candidate items.

A dataset for this tasks would probably be feeding your model with dictionaries somewhat like this:

{"query_id": [10, 1234, 23470],
 "candidate_id": [12380, 871029, 7141]

[max-grzanna]

Thanks, @maciejkula, for answering my question. That really helps!
I think I am going with a shared table. If I want to use context features like the item's manufacturer or color, should I add them to the query or candidate tower for item-to-item recommendation? Or maybe to both?

[maciejkula]

Both seems like a good starting point!

[max-grzanna]

Okay, thanks! I've wondered if you know where to find some sample code on item-to-item recommendation with TFRS?

[luisceledon]

Hi @max-grzanna , did you find any code example for this? thanks in advance

[max-grzanna]

No I haven’t :/ Sorry.

[piyushdewangan54]

Hi @max-grzanna, have got any code regarding this TFRS ??
Do you have any idea how can we create a shared table for Amazon review dataset ??