Feature/cold warm for vector models (#122)

- Added `dataset` argument to `recommend_for_cold` method
- Supported warm and cold reco in `LightFMWrapperModel`
- Supported warm reco in `DSSMModels` and in corresponding pytorch
datasets
- [Breaking] Replaced `include_warm` parameter in
`Dataset.get_user_item_matrix` to pair `include_warm_users` and
`include_warm_items`
- [Breaking] Added `n_factors` and `deterministic` params to `DSSMModel`
- [Breaking] Added default value for train dataset type and params for
user and item dataset types in `DSSMModel`
- [Breaking] Replaced `include_warm` parameter in
`Dataset.get_user_item_matrix` to pair `include_warm_users` and
`include_warm_items`
- [Breaking] Renamed torch datasets and `dataset_type` to
`train_dataset_type` param in `DSSMModel`

Author: feldlime

.pylintrc

@@ -441,7 +441,7 @@ valid-metaclass-classmethod-first-arg=cls
 [DESIGN]
 
 # Maximum number of arguments for function / method.
-max-args=15
+max-args=20
 
 # Maximum number of attributes for a class (see R0902).
 max-attributes=12

CHANGELOG.md

@@ -10,12 +10,17 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ### Added 
 - Warm users/items support in `Dataset` ([#77](https://github.com/MobileTeleSystems/RecTools/pull/77))
-- Warm and cold users/items support in `ModelBase` and non-personalized models ([#77](https://github.com/MobileTeleSystems/RecTools/pull/77), [#120](https://github.com/MobileTeleSystems/RecTools/pull/120))
+- Warm and cold users/items support in `ModelBase` and all possible models ([#77](https://github.com/MobileTeleSystems/RecTools/pull/77), [#120](https://github.com/MobileTeleSystems/RecTools/pull/120), [#122](https://github.com/MobileTeleSystems/RecTools/pull/122))
 - Warm and cold users/items support in `cross_validate` ([#77](https://github.com/MobileTeleSystems/RecTools/pull/77))
+- [Breaking] Default value for train dataset type and params for user and item dataset types in `DSSMModel` ([#122](https://github.com/MobileTeleSystems/RecTools/pull/122))
+- [Breaking] `n_factors` and `deterministic` params to `DSSMModel` ([#122](https://github.com/MobileTeleSystems/RecTools/pull/122))
 
 ### Changed
 - Changed the logic of choosing random sampler for `RandomModel` and increased the sampling speed ([#120](https://github.com/MobileTeleSystems/RecTools/pull/120))
-- Changed the logic of `RandomModel`: now the recommendations are different for repeated calls of recommend methods ([#120](https://github.com/MobileTeleSystems/RecTools/pull/120))
+- [Breaking] Changed the logic of `RandomModel`: now the recommendations are different for repeated calls of recommend methods ([#120](https://github.com/MobileTeleSystems/RecTools/pull/120))
+- Torch datasets to support warm recommendations ([#122](https://github.com/MobileTeleSystems/RecTools/pull/122))
+- [Breaking] Replaced `include_warm` parameter in `Dataset.get_user_item_matrix` to pair `include_warm_users` and `include_warm_items` ([#122](https://github.com/MobileTeleSystems/RecTools/pull/122))
+- [Breaking] Renamed torch datasets and `dataset_type` to `train_dataset_type` param in `DSSMModel` ([#122](https://github.com/MobileTeleSystems/RecTools/pull/122))
 
 ### Removed
 - `return_external_ids` parameter in `recommend` and `recommend_to_items` model methods ([#77](https://github.com/MobileTeleSystems/RecTools/pull/77))

rectools/dataset/dataset.py

@@ -194,7 +194,12 @@ def _make_features(
         except Exception as e:  # pragma: no cover
             raise RuntimeError(f"An error has occurred while constructing {feature_type} features: {e!r}")
 
-    def get_user_item_matrix(self, include_weights: bool = True, include_warm: bool = False) -> sparse.csr_matrix:
+    def get_user_item_matrix(
+        self,
+        include_weights: bool = True,
+        include_warm_users: bool = False,
+        include_warm_items: bool = False,
+    ) -> sparse.csr_matrix:
         """
         Construct user-item CSR matrix based on `interactions` attribute.
 
@@ -219,8 +224,9 @@ def get_user_item_matrix(self, include_weights: bool = True, include_warm: bool
             Resized user-item CSR matrix
         """
         matrix = self.interactions.get_user_item_matrix(include_weights)
-        if include_warm:
-            matrix.resize(self.user_id_map.size, self.item_id_map.size)
+        n_rows = self.user_id_map.size if include_warm_users else matrix.shape[0]
+        n_columns = self.item_id_map.size if include_warm_items else matrix.shape[1]
+        matrix.resize(n_rows, n_columns)
         return matrix
 
     def get_raw_interactions(self, include_weight: bool = True, include_datetime: bool = True) -> pd.DataFrame:

rectools/dataset/torch_datasets.py

@@ -25,12 +25,23 @@
 
 from .dataset import Dataset
 
-DD = tp.TypeVar("DD", bound="DSSMDataset")
-ID = tp.TypeVar("ID", bound="ItemFeaturesDataset")
-UD = tp.TypeVar("UD", bound="UserFeaturesDataset")
+DSSMTrainDatasetT = tp.TypeVar("DSSMTrainDatasetT", bound="DSSMTrainDatasetBase")
+DSSMItemDatasetT = tp.TypeVar("DSSMItemDatasetT", bound="DSSMItemDatasetBase")
+DSSMUserDatasetT = tp.TypeVar("DSSMUserDatasetT", bound="DSSMUserDatasetBase")
 
 
-class DSSMDataset(TorchDataset[tp.Any]):
+class DSSMTrainDatasetBase(TorchDataset[tp.Any]):
+    """Base class for DSSM training datasets. Used only for type hinting."""
+
+    def __init__(self, *args: tp.Any, **kwargs: tp.Any) -> None:
+        raise NotImplementedError()
+
+    @classmethod
+    def from_dataset(cls: tp.Type[DSSMTrainDatasetT], dataset: Dataset) -> DSSMTrainDatasetT:
+        raise NotImplementedError()
+
+
+class DSSMTrainDataset(DSSMTrainDatasetBase):
     """
     Torch dataset wrapper for `rectools.dataset.dataset.Dataset`.
     Implements `torch.utils.data.Dataset` for subsequent usage with
@@ -68,7 +79,7 @@ def __init__(
             )
 
     @classmethod
-    def from_dataset(cls: tp.Type[DD], dataset: Dataset) -> DD:
+    def from_dataset(cls: tp.Type[DSSMTrainDatasetT], dataset: Dataset) -> DSSMTrainDatasetT:
         ui_matrix = dataset.get_user_item_matrix()
 
         # We take hot here since this dataset is used for fit only
@@ -99,7 +110,18 @@ def __getitem__(
         return user_features, interactions, pos, neg
 
 
-class ItemFeaturesDataset(TorchDataset[tp.Any]):
+class DSSMItemDatasetBase(TorchDataset[tp.Any]):
+    """Base class for DSSM item datasets. Used only for type hinting."""
+
+    def __init__(self, *args: tp.Any, **kwargs: tp.Any) -> None:
+        raise NotImplementedError()
+
+    @classmethod
+    def from_dataset(cls: tp.Type[DSSMItemDatasetT], dataset: Dataset) -> DSSMItemDatasetT:
+        raise NotImplementedError()
+
+
+class DSSMItemDataset(DSSMItemDatasetBase):
     """
     Torch dataset wrapper for `rectools.dataset.dataset.Dataset`.
     Implements `torch.utils.data.Dataset` for subsequent usage with
@@ -113,7 +135,7 @@ def __init__(self, items: sparse.csr_matrix):
         self.items = items
 
     @classmethod
-    def from_dataset(cls: tp.Type[ID], dataset: Dataset) -> ID:
+    def from_dataset(cls: tp.Type[DSSMItemDatasetT], dataset: Dataset) -> DSSMItemDatasetT:
         # We take all features here since this dataset is used for recommend only, not for fit
         if dataset.item_features is not None:
             return cls(dataset.item_features.get_sparse())
@@ -126,7 +148,22 @@ def __getitem__(self, idx: int) -> torch.FloatTensor:
         return torch.FloatTensor(self.items[idx].toarray().flatten())
 
 
-class UserFeaturesDataset(TorchDataset[tp.Any]):
+class DSSMUserDatasetBase(TorchDataset[tp.Any]):
+    """Base class for DSSM training datasets. Used only for type hinting."""
+
+    def __init__(self, *args: tp.Any, **kwargs: tp.Any) -> None:
+        raise NotImplementedError()
+
+    @classmethod
+    def from_dataset(
+        cls: tp.Type[DSSMUserDatasetT],
+        dataset: Dataset,
+        keep_users: tp.Optional[tp.Sequence[int]] = None,
+    ) -> DSSMUserDatasetT:
+        raise NotImplementedError()
+
+
+class DSSMUserDataset(DSSMUserDatasetBase):
     """
     Torch dataset wrapper for `rectools.dataset.dataset.Dataset`.
     Implements `torch.utils.data.Dataset` for subsequent usage with
@@ -143,6 +180,8 @@ def __init__(
         interactions: sparse.csr_matrix,
         keep_users: tp.Optional[tp.Sequence[int]] = None,
     ):
+        if users.shape[0] != interactions.shape[0]:
+            raise ValueError("Number of rows in user features matrix and in interactions matrix must be the same")
         if keep_users is not None:
             self.users = users[keep_users]
             self.interactions = interactions[keep_users]
@@ -152,15 +191,15 @@ def __init__(
 
     @classmethod
     def from_dataset(
-        cls: tp.Type[UD],
+        cls: tp.Type[DSSMUserDatasetT],
         dataset: Dataset,
         keep_users: tp.Optional[tp.Sequence[int]] = None,
-    ) -> UD:
+    ) -> DSSMUserDatasetT:
         # We take all features here since this dataset is used for recommend only, not for fit
         if dataset.user_features is not None:
             return cls(
                 dataset.user_features.get_sparse(),
-                dataset.get_user_item_matrix(),
+                dataset.get_user_item_matrix(include_warm_users=True),
                 keep_users,
             )
         raise AttributeError("User features attribute of dataset could not be None")

rectools/models/base.py

@@ -166,9 +166,11 @@ def recommend(
                 reco_warm = self._recommend_u2i_warm(warm_user_ids, dataset, k, sorted_item_ids_to_recommend)
             else:
                 # TODO: use correct types for numpy arrays and stop ignoring
-                reco_warm = self._recommend_cold(warm_user_ids, k, sorted_item_ids_to_recommend)  # type: ignore
+                reco_warm = self._recommend_cold(
+                    warm_user_ids, dataset, k, sorted_item_ids_to_recommend
+                )  # type: ignore
         if cold_user_ids.size > 0:
-            reco_cold = self._recommend_cold(cold_user_ids, k, sorted_item_ids_to_recommend)
+            reco_cold = self._recommend_cold(cold_user_ids, dataset, k, sorted_item_ids_to_recommend)
 
         reco_hot = self._adjust_reco_types(reco_hot)
         reco_warm = self._adjust_reco_types(reco_warm)
@@ -285,11 +287,11 @@ def recommend_to_items(  # pylint: disable=too-many-branches
             else:
                 # TODO: use correct types for numpy arrays and stop ignoring
                 reco_warm = self._recommend_cold(
-                    warm_target_ids, requested_k, sorted_item_ids_to_recommend
+                    warm_target_ids, dataset, requested_k, sorted_item_ids_to_recommend
                 )  # type: ignore
         if cold_target_ids.size > 0:
             # We intentionally request `k` and not `requested_k` here since we're not going to filter cold reco later
-            reco_cold = self._recommend_cold(cold_target_ids, k, sorted_item_ids_to_recommend)
+            reco_cold = self._recommend_cold(cold_target_ids, dataset, k, sorted_item_ids_to_recommend)
 
         reco_hot = self._adjust_reco_types(reco_hot)
         reco_warm = self._adjust_reco_types(reco_warm)
@@ -459,7 +461,11 @@ def _make_reco_table(cls, reco: RecoTriplet, target_col: str, add_rank_col: bool
         return df
 
     def _recommend_cold(
-        self, target_ids: AnyIdsArray, k: int, sorted_item_ids_to_recommend: tp.Optional[InternalIdsArray]
+        self,
+        target_ids: AnyIdsArray,
+        dataset: Dataset,
+        k: int,
+        sorted_item_ids_to_recommend: tp.Optional[InternalIdsArray],
     ) -> SemiInternalRecoTriplet:
         raise NotImplementedError()
 
@@ -509,16 +515,20 @@ class FixedColdRecoModelMixin:
     """
 
     def _recommend_cold(
-        self, target_ids: AnyIdsArray, k: int, sorted_item_ids_to_recommend: tp.Optional[InternalIdsArray]
+        self,
+        target_ids: AnyIdsArray,
+        dataset: Dataset,
+        k: int,
+        sorted_item_ids_to_recommend: tp.Optional[InternalIdsArray],
     ) -> SemiInternalRecoTriplet:
-        item_ids, scores = self._get_cold_reco(k, sorted_item_ids_to_recommend)
+        item_ids, scores = self._get_cold_reco(dataset, k, sorted_item_ids_to_recommend)
         reco_target_ids = np.repeat(target_ids, len(item_ids))
         reco_item_ids = np.tile(item_ids, len(target_ids))
         reco_scores = np.tile(scores, len(target_ids))
 
         return reco_target_ids, reco_item_ids, reco_scores
 
     def _get_cold_reco(
-        self, k: int, sorted_item_ids_to_recommend: tp.Optional[InternalIdsArray]
+        self, dataset: Dataset, k: int, sorted_item_ids_to_recommend: tp.Optional[InternalIdsArray]
     ) -> tp.Tuple[InternalIds, Scores]:
         raise NotImplementedError()