How to compute rank_items for all my users?  #384

Activity

1. mohjeih commented on Aug 4, 2020

   

   mohjeih

   on Aug 4, 2020

   More actions

   Hello,

   One solution could be implementing your own rank_item_all method with a custom C++ function filling numpy arrays in parallel, similar to recommend_all method.
2. ManuelRios18 commented on Aug 4, 2020

   

   ManuelRios18

   on Aug 4, 2020

   Author

   More actions

   Hello,

   One solution could be implementing your own rank_item_all method with a custom C++ function filling numpy arrays in parallel, similar to recommend_all method.

   @mohjeih
   Where can I get the recommend_all method written in C++ ? And where is the current "rank_items" method implemented ?
   In order to have a guideline
3. mohjeih commented on Aug 4, 2020

   

   mohjeih

   on Aug 4, 2020

   More actions

   @ManuelRios18 Sorry for the confusion. recommend_all and rank_items are implemented in recommender_base.pyx (Cythonized module). Please see here .recommend_all sorts array of scores in parallel using a C++ sorting function, fargsort_c, which is implemented in topnc.cpp. The wrapper to this sorting function is defined at the beginning of recommender_base.pyx.
4. marinkreso95 commented on Apr 9, 2021

   

   marinkreso95

   on Apr 9, 2021

   Last edited by marinkreso95

   More actions

   Hi
   I have the same problem as @ManuelRios18 (if you found a solution for this, please share)
   Here is a code snippet from my project. Does somebody have another way to do this properly? Because this solution is taking a lot of time

       row = y.recipe_id_comb.astype(item_c).cat.codes
       col = y.user_id.astype(user_c).cat.codes
       
       item_user_matrix = csr_matrix((y["rating"], (row, col)), shape=(item_c.categories.size, user_c.categories.size))
       user_item_matrix = csr_matrix((y["rating"], (col, row)), shape=(user_c.categories.size, item_c.categories.size))
       
       # Train MF model
       mf_model = implicit.als.AlternatingLeastSquares(factors=opt_fac, regularization=opt_reg, iterations=opt_itr)
       mf_model.fit(item_user_matrix)
   
       # Extract outputs from MF model
       score_mf = row.copy()
       for ind_i in np.unique(col):
           mf_y = mf_model.rank_items(ind_i, user_item_matrix, row[col == ind_i].astype(int))
           score_mf[col == ind_i] = score_mf[col == ind_i].replace(dict(mf_y))
       score_mf = score_mf.to_numpy().reshape((len(score_mf), 1))
   
5. ManuelRios18 commented on Apr 10, 2021

   

   ManuelRios18

   on Apr 10, 2021

   Author

   More actions

   Hello, Sadly I couldn't find an efficient solution, what I ended up doing was simply multiply user and item embeddings, load those results into my company's Landing Zone, and then I ranked the items using basic SQL queries.

   Happy coding!

   @marinkreso95
6. benfred commented on Jan 3, 2022

   

   benfred

   on Jan 3, 2022

   Owner

   More actions

   @ManuelRios18 the latest main branch has support for this (though does have some breaking API changes ). see #481

   model = implicit.als.AlternatingLeastSquares()
   model.fit(user_items)
   
   # rank the first 200 items in the model for all the users
   userids = np.arange(model.user_factors.shape[0])
   items = np.arange(200)
   ids, scores = model.recommend(userids, items=items)
7. benfred

   closed this as completedon Jan 3, 2022
8. benfred

   linked a pull request that will close this issueApi refactor #494on Jan 3, 2022