API Refactor #481

The current API could be improved:

• Return numpy arrays from recommend methods #482
  We currently return recommend results as python lists of tuples of (itemid, score) . We should instead return a numpy array of the itemids, and another numpy array of scores.

• Change model.fit to take a user_items sparse matrix #484
  model.fit takes an item_users sparse matrix, while model.recommend takes a user_items sparse matrix. It would cause less confusion if both of these methods took a user_items matrix instead. The ALS, LMF and KNN models need both user_items and item_users - so calculate the transpose as the first step, meaning it doesn't really matter which is passed. The BPR model only needs user_items, so would be better to have a user_items matrix.

• Unify API for rank_items/recommend/recommend_all #489
  We have 3 different methods for recommending items to users : recommend, recommend_all and rank_items. All three of these are inconsistent with the others:

  • recommend_all doesn’t return scores for recommendations, just ranked items - whereas recommend/rank_items do
  • recommend_all throws errors if we would filter out too many items for any user, where rank_items/recommend just return fewer items
  • rank_items doesn’t support filtering users own likes, whereas recommend/recommend_all do
  • recommend has a batch mode operation in recommend_all - but rank_items doesn't have a corresponding batch mode
  • recommend_all only works on the CPU versions of MF models - and isn't included on KNN or GPU models

  I think we should consolidate down to a single method (recommend) which will take either a single userid or a vector of userids to compute recommendations for. We should augment the filtering options, which currently allow you to exclude items, with another parameter (items) which will force only including items on it. This is like adding an item whitelist to the recommend method, to complement the filter_items blacklist parameter we already have - and would make rank_items redundant.

• Filtering for similar_items and similar_items #488
  The similar_items and similar_users methods don't have the equivalent filtering options that are included in the recommend method. We should also provide the same filtering options for similar_items and similar_users.

• Approximate nearest neighbour for BPR/LMF and GPU models #487
  Approximate nearest neighbours only works for the ALS model - and not on BPR or LMF. Likewise this doesn't work on the GPU models. We should change the ANN models to be composable - and wrap any other MF model

• Don't require empty rows in user_items and item_users #526

Where possible we should add deprecation warnings and maintain the legacy API for a couple releases. The first two changes will be unavoidably breaking API changes.