predict_rank is slow for all items of a single user

[DeoLeung]

I have a model of (30k, 1m) user/item
when I use predict_rank to predict all the ranks of a single user, it gets slow:

interactions = csr_matrix(
          ([1] * item_size, ([user_id] * item_size, item_ids)),
          shape=(user_size, item_size),
          dtype=np.float32)
# this takes very long
model.predict_rank(interactions)
# very fast
model.predict(user_id, item_ids)

isn't the ranks are just the recommended scores' order?
I guess predict_rank do predict for every interaction？

thanks,

[andyluther]

i believe predict_rank predicts the rank of every item for each user in interactions

[maciejkula]

This method is to be used when you're only evaluating a couple of interactions for every user, as is most common in evaluating models. It is quadratic in the number of nonzero interactions per user, so if you want to evaluate more items than that I suggest you use predict instead.

I updated the docstring to reflect this in #111

[adsk2050]

This method is to be used when you're only evaluating a couple of interactions for every user, as is most common in evaluating models. It is quadratic in the number of nonzero interactions per user, so if you want to evaluate more items than that I suggest you use predict instead.

I updated the docstring to reflect this in #111

In that case shouldn't the precision_at_k code be made to run on predict instead of predict_rank/predict_ranks? I am asking because precision_at_k is also very slow.

[tinawenzel]

This method is to be used when you're only evaluating a couple of interactions for every user, as is most common in evaluating models. It is quadratic in the number of nonzero interactions per user, so if you want to evaluate more items than that I suggest you use predict instead.
I updated the docstring to reflect this in #111

In that case shouldn't the precision_at_k code be made to run on predict instead of predict_rank/predict_ranks? I am asking because precision_at_k is also very slow.

@adsk2050 I agree. I tried to replicate prec@k using predict, but using predict results in a lot smaller prec@k scores than predict_rank. See #568.