Update content personalization

docs/releases/unreleased.md

@@ -13,3 +13,8 @@ River's mini-batch methods now support pandas v2. In particular, River conforms
 ## forest
 
 - Simplify inner the structures of `forest.ARFClassifier` and `forest.ARFRegressor` by removing redundant class hierarchy. Simplify how concept drift logging can be accessed in individual trees and in the forest as a whole.
+
+## reco
+
+- Renamed `reco.base.Ranker` to `reco.base.Recommender`.
+- Added a `sample(user, items)` method to `reco.base.Recommender`.

river/bandit/lin_ucb.py

@@ -36,7 +36,7 @@ class LinUCBDisjoint(bandit.base.ContextualPolicy):
     References
     ----------
     [^1]: [A Contextual-Bandit Approach to Personalized News Article Recommendation](https://arxiv.org/abs/1003.0146)
-    [^2:] [Contextual Bandits Analysis of LinUCB Disjoint Algorithm with Dataset](https://kfoofw.github.io/contextual-bandits-linear-ucb-disjoint/)
+    [^2]: [Contextual Bandits Analysis of LinUCB Disjoint Algorithm with Dataset](https://kfoofw.github.io/contextual-bandits-linear-ucb-disjoint/)
 
     """
 

river/checks/__init__.py

@@ -8,7 +8,7 @@
 
 from river.base import Estimator
 from river.model_selection.base import ModelSelector
-from river.reco.base import Ranker
+from river.reco.base import Recommender
 
 from . import anomaly, clf, common, model_selection, reco
 
@@ -54,7 +54,7 @@ def _yield_datasets(model: Estimator):
 
     # Recommendation models can be regressors or classifiers, but they have requirements as to the
     # structure of the data
-    if isinstance(utils.inspect.extract_relevant(model), Ranker):
+    if isinstance(utils.inspect.extract_relevant(model), Recommender):
         if utils.inspect.isregressor(model):
             yield _DummyDataset(
                 ("Alice", "Superman", 8),
@@ -164,7 +164,7 @@ def yield_checks(model: Estimator) -> typing.Iterator[typing.Callable]:
     if isinstance(utils.inspect.extract_relevant(model), ModelSelector):
         dataset_checks.append(model_selection.check_model_selection_order_does_not_matter)
 
-    if isinstance(utils.inspect.extract_relevant(model), Ranker):
+    if isinstance(utils.inspect.extract_relevant(model), Recommender):
         yield reco.check_reco_routine
 
     if utils.inspect.isanomalydetector(model):

river/reco/base.py

@@ -11,10 +11,10 @@
 Reward = typing.Union[numbers.Number, bool]  # noqa: UP007
 
 
-__all__ = ["Ranker"]
+__all__ = ["Recommender"]
 
 
-class Ranker(base.Estimator):
+class Recommender(base.Estimator):
     """Base class for ranking models.
 
     Parameters
@@ -68,8 +68,24 @@ def predict_one(self, user: ID, item: ID, x: dict | None = None) -> Reward:
 
         """
 
+    def sample(self, user: ID, items: set[ID], x: dict | None = None):
+        """Sample an item at random, based on the preference of a given user.
+
+        Parameters
+        ----------
+        user
+            A user ID.
+        items
+            A set of items to rank.
+        x
+            Optional context to use.
+
+        """
+        preferences = [max(1e-10, self.predict_one(user, item, x)) for item in items]
+        return self._rng.choices(items, weights=preferences, k=1)[0]
+
     def rank(self, user: ID, items: set[ID], x: dict | None = None) -> list[ID]:
-        """Rank models by decreasing order of preference for a given user.
+        """Rank items by decreasing order of preference for a given user.
 
         Parameters
         ----------

river/reco/baseline.py

@@ -8,7 +8,7 @@
 __all__ = ["Baseline"]
 
 
-class Baseline(reco.base.Ranker):
+class Baseline(reco.base.Recommender):
     """Baseline for recommender systems.
 
     A first-order approximation of the bias involved in target. The model equation is defined as:

river/reco/biased_mf.py

@@ -6,14 +6,12 @@
 
 import numpy as np
 
-from river import optim, stats, utils
-
-from .base import Ranker
+from river import optim, reco, stats, utils
 
 __all__ = ["BiasedMF"]
 
 
-class BiasedMF(Ranker):
+class BiasedMF(reco.base.Recommender):
     """Biased Matrix Factorization for recommender systems.
 
     The model equation is defined as:

river/reco/funk_mf.py

@@ -11,7 +11,7 @@
 __all__ = ["FunkMF"]
 
 
-class FunkMF(reco.base.Ranker):
+class FunkMF(reco.base.Recommender):
     """Funk Matrix Factorization for recommender systems.
 
     The model equation is defined as:
@@ -131,7 +131,7 @@ def predict_one(self, user, item, x=None):
 
     def learn_one(self, user, item, y, x=None):
         # Calculate the gradient of the loss with respect to the prediction
-        g_loss = self.loss.gradient(y, self.predict_one(user, item))
+        g_loss = self.loss.gradient(y, self.predict_one(user, item, x=x))
 
         # Clamp the gradient to avoid numerical instability
         g_loss = utils.math.clamp(g_loss, minimum=-self.clip_gradient, maximum=self.clip_gradient)

river/reco/normal.py

@@ -5,7 +5,7 @@
 __all__ = ["RandomNormal"]
 
 
-class RandomNormal(reco.base.Ranker):
+class RandomNormal(reco.base.Recommender):
     """Predicts random values sampled from a normal distribution.
 
     The parameters of the normal distribution are fitted with running statistics. They parameters