Merge pull request #139 from ita9naiwa/add-similar-users

Add similar users calculation in MF models

implicit/nearest_neighbours.py

@@ -62,6 +62,9 @@ def rank_items(self, userid, user_items, selected_items, recalculate_user=False)
                 ret.append((itemid, -1.0))
         return ret
 
+    def similar_users(self, userid, N=10):
+        raise NotImplementedError("Not implemented Yet")
+
     def similar_items(self, itemid, N=10):
         """ Returns a list of the most similar other items """
         if itemid >= self.similarity.shape[0]:

implicit/recommender_base.py

@@ -83,6 +83,25 @@ def rank_items(self, userid, user_items, selected_items, recalculate_user=False)
         """
         pass
 
+    @abstractmethod
+    def similar_users(self, userid, N=10):
+        """
+        Calculates a list of similar items
+
+        Parameters
+        ----------
+        userid : int
+            The row id of the user to retrieve similar users for
+        N : int, optional
+            The number of similar users to return
+
+        Returns
+        -------
+        list
+            List of (userid, score) tuples
+        """
+        pass
+
     @abstractmethod
     def similar_items(self, itemid, N=10):
         """
@@ -118,8 +137,8 @@ def __init__(self):
         self.item_factors = None
         self.user_factors = None
 
-        # cache of item norms (useful for calculating similar items)
-        self._item_norms = None
+        # cache of user, item norms (useful for calculating similar items)
+        self._user_norms, self._item_norms = None, None
 
     def recommend(self, userid, user_items,
                   N=10, filter_already_liked_items=True, filter_items=None, recalculate_user=False):
@@ -167,13 +186,37 @@ def _user_factor(self, userid, user_items, recalculate_user=False):
     def recalculate_user(self, userid, user_items):
         raise NotImplementedError("recalculate_user is not supported with this model")
 
+    def similar_users(self, userid, N=10):
+        factor = self.user_factors[userid]
+        factors = self.user_factors
+        norms = self.user_norms
+
+        return self._get_similarity_score(factor, factors, norms, N)
+
+    similar_users.__doc__ = RecommenderBase.similar_users.__doc__
+
     def similar_items(self, itemid, N=10):
-        scores = self.item_factors.dot(self.item_factors[itemid]) / self.item_norms
-        best = np.argpartition(scores, -N)[-N:]
-        return sorted(zip(best, scores[best] / self.item_norms[itemid]), key=lambda x: -x[1])
+        factor = self.item_factors[itemid]
+        factors = self.item_factors
+        norms = self.item_norms
+
+        return self._get_similarity_score(factor, factors, norms, N)
 
     similar_items.__doc__ = RecommenderBase.similar_items.__doc__
 
+    def _get_similarity_score(self, factor, factors, norms, N):
+        scores = factors.dot(factor) / norms
+        best = np.argpartition(scores, -N)[-N:]
+        return sorted(zip(best, scores[best]), key=lambda x: -x[1])
+
+    @property
+    def user_norms(self):
+        if self._user_norms is None:
+            self._user_norms = np.linalg.norm(self.user_factors, axis=-1)
+            # don't divide by zero in similar_items, replace with small value
+            self._user_norms[self._user_norms == 0] = 1e-10
+        return self._user_norms
+
     @property
     def item_norms(self):
         if self._item_norms is None:

tests/recommender_base_test.py

@@ -6,6 +6,7 @@
 from scipy.sparse import csr_matrix
 
 from implicit.evaluation import precision_at_k
+from implicit.nearest_neighbours import ItemItemRecommender
 
 
 class TestRecommenderBaseMixin(object):
@@ -69,6 +70,19 @@ def test_evaluation(self):
                            show_progress=False)
         self.assertEqual(p, 1)
 
+    def test_similar_users(self):
+
+        model = self._get_model()
+        # calculating similar users in nearest-neighbours is not implemented yet
+        if isinstance(model, ItemItemRecommender):
+            return
+        model.show_progress = False
+        model.fit(self.get_checker_board(50))
+        for userid in range(50):
+            recs = model.similar_users(userid, N=10)
+            for r, _ in recs:
+                self.assertEqual(r % 2, userid % 2)
+
     def test_similar_items(self):
         model = self._get_model()
         model.show_progress = False