Standardize predict interface using SAR standard

[miguelgfierro]

Description

for rating: predict
for ranking: recommend_top_k
also applied black

Related Issues

Checklist:

• I have followed the contribution guidelines and code style for this project.
• I have added tests covering my contributions.
• I have updated the documentation accordingly.
• This PR is being made to staging and not master.

[review-notebook-app]

Check out this pull request on 

You'll be able to see Jupyter notebook diff and discuss changes. Powered by ReviewNB.

[anargyri]

This function does not recommend k items. It computes the predictions for all users and items.

[miguelgfierro]

ohh good catch, the cut of k is done in fact when the metric is computed https://github.com/microsoft/recommenders/blob/master/notebooks/02_model/surprise_svd_deep_dive.ipynb, the output is a massive matrix instead of n_users x k

[miguelgfierro]

after the meeting, we need to check whether the compute_ranking_pred method is used in another function that needs the full matrix, if not, we can optimize it and

for user in data[usercol].unique():
        for item in data[itemcol].unique():
            preds_lst.append(algo.predict(user, item).est)
       preds = sort(preds_lst)
       preds = preds[:k]
       
       preds_lst.append(preds)

[miguelgfierro]

also move this up:

    if remove_seen:
        tempdf = pd.concat(
            [
                data[[usercol, itemcol]],
                pd.DataFrame(
                    data=np.ones(data.shape[0]), columns=["dummycol"], index=data.index
                ),
            ],
            axis=1,
        )
        merged = pd.merge(tempdf, all_predictions, on=[usercol, itemcol], how="outer")
        return merged[merged["dummycol"].isnull()].drop("dummycol", axis=1)
    else:
        return all_predictions

so we remove the seen items before sorting

[miguelgfierro]

also check with @yueguoguo whether or not we are using the threshold cut #1041

[gramhagen]

if we don't want to work with the scores as a matrix like sar does we could use a heap

from heapq import heappush, heappushpop

users = data[usercol].unique()
items = data[itemcol].unique()
preds_lst = np.zeros([len(users), k])

for user_idx, user in enumerate(users):
    heap = []
    for item in items:
        score = algo.predict(user, item).est
        if len(heap) < k:
            heappush(heap, score)
        elif score > heap[k - 1]:
            heappushpop(heap, score);
    preds_lst[user_idx] = heap[::-1]

[anargyri]

heapq also has an nlargest() method. Not sure which way would be faster (depends on which algorithm it uses).

[gramhagen]

true, but it wasn't clear to me if that method tries to sort the items (which would be unnecessary) or leverages the heap?

[anargyri]

https://github.com/python/cpython/blob/61b3484cdf27ceca1c1069a351487d2db4b2b48c/Lib/heapq.py#L395
It looks similar to your for loop.

[miguelgfierro]

hey guys @anargyri @gramhagen, I reverted the interface of ranking in surprise to do the work on a different PR. The related issue is here: #1042.

Please let me know if you see anything else on this PR