[Question] What are best practices to post-process predictions for inventory levels?

[ageofneil]

Suppose I run the worlds last DVD-rent-by-mail service and we've developed a basic recommend model by following your excellent Recommending movies: retrieval guide. Everything is working great until we notice our model is recommending movies that are out of stock! What's a good way to deal with this? I can't train the model before every recommendation, so obviously some sort of post processing needs to be done to the recommendation. I'm not sure the right way to approach this.

[dartpain]

For example for

index= tfrs.layers.factorized_top_k.ScaNN(model.query_model, num_reordering_candidates=1000)
index.index_from_dataset(
    tf.data.Dataset.zip((items_map.batch(100), 
    items.batch(100).map(model.candidate_model)))
)

To query it and get recomendations with only in stock items you can create a list with out of stock items, items_exclusion_list

and use it like this

x, t = index.query_with_exclusions(query_doc, exclusions=tf.constant([items_exclusion_list]))

Otherwise,

You can get large number of recomendations and remove out of stock items

[maciejkula]

@dartpain makes a great suggestion.

I will also point out that recommenders are almost always deployed within a larger system that contains a final business logic layer - in this case, that business logic layer would make database calls to ensure that whatever is returned to the user is in stock.