Removed internal ids usage outside models (#177)

- Removed `assume_external_ids` parameter in `recommend` and
`recommend_to_items` model methods
- Added `on_unsupported_targets` in `recommend` and `recommend_to_items`
model methods
- Added `filter_on_interactions_df_row_indexes` method of `Dataset`
- Fixed `IntraListDiversity` metric computation in `cross_validate`

Author: blondered

CHANGELOG.md

@@ -11,13 +11,17 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 ### Added
 - `Debias` mechanism for classification, ranking and auc metrics. New parameter `is_debiased` to `calc_from_confusion_df`, `calc_per_user_from_confusion_df` methods of classification metrics, `calc_from_fitted`, `calc_per_user_from_fitted` methods of auc and rankning (`MAP`) metrics, `calc_from_merged`, `calc_per_user_from_merged` methods of ranking (`NDCG`, `MRR`) metrics. ([#152](https://github.com/MobileTeleSystems/RecTools/pull/152))
 - `nbformat >= 4.2.0` dependency to `[visuals]` extra ([#169](https://github.com/MobileTeleSystems/RecTools/pull/169))
+- `filter_interactions` method of `Dataset` ([#177](https://github.com/MobileTeleSystems/RecTools/pull/177))
+- `on_unsupported_targets` parameter to  `recommend` and `recommend_to_items` model methods ([#177](https://github.com/MobileTeleSystems/RecTools/pull/177))
 
 ### Fixed
 - `display()` method in `MetricsApp` ([#169](https://github.com/MobileTeleSystems/RecTools/pull/169))
-
-### Fixed
+- `IntraListDiversity` metric computation in `cross_validate` ([#177](https://github.com/MobileTeleSystems/RecTools/pull/177))
 - Allow warp-kos loss for LightFMWrapperModel ([#175](https://github.com/MobileTeleSystems/RecTools/pull/175))
 
+### Removed
+- [Breaking] `assume_external_ids` parameter in `recommend` and `recommend_to_items` model methods ([#177](https://github.com/MobileTeleSystems/RecTools/pull/177))
+
 ## [0.7.0] - 29.07.2024
 
 ### Added

README.md

@@ -79,7 +79,7 @@ RecTools is on PyPI, so you can use `pip` to install it.
 ```
 pip install rectools
 ```
-The default version doesn't contain all the dependencies, because some of them are needed only for specific models. Available user extensions are the following:
+The default version doesn't contain all the dependencies, because some of them are needed only for specific functionality. Available user extensions are the following:
 
 - `lightfm`: adds wrapper for LightFM model,
 - `torch`: adds models based on neural nets,

rectools/dataset/dataset.py

@@ -17,6 +17,7 @@
 import typing as tp
 
 import attr
+import numpy as np
 import pandas as pd
 from scipy import sparse
 
@@ -245,3 +246,66 @@ def get_raw_interactions(self, include_weight: bool = True, include_datetime: bo
         pd.DataFrame
         """
         return self.interactions.to_external(self.user_id_map, self.item_id_map, include_weight, include_datetime)
+
+    def filter_interactions(
+        self,
+        row_indexes_to_keep: np.ndarray,
+        keep_external_ids: bool = True,
+        keep_features_for_removed_entities: bool = True,
+    ) -> "Dataset":
+        """
+        Generate filtered dataset that contains only provided `row_indexes_to_keep` from original
+        dataset interactions dataframe.
+        Resulting dataset will get new id mapping for both users and items.
+
+        Parameters
+        ----------
+        row_indexes_to_keep : np.ndarray
+            Original dataset interactions df row indexes that are to be kept
+        keep_external_ids : bool, default `True`
+            Whether to keep external ids -> 2x internal ids mapping (default).
+            Otherwise internal -> 2x internal ids mapping will be created.
+        keep_features_for_removed_entities : bool, default `True`
+            Whether to keep all features for users and items that are not hot any more.
+
+        Returns
+        -------
+        Dataset
+            Filtered dataset that has only selected interactions, new ids mapping and processed features.
+        """
+        interactions_df = self.interactions.df.iloc[row_indexes_to_keep]
+
+        # 1x internal -> 2x internal
+        user_id_map = IdMap.from_values(interactions_df[Columns.User].values)
+        item_id_map = IdMap.from_values(interactions_df[Columns.Item].values)
+        interactions = Interactions.from_raw(interactions_df, user_id_map, item_id_map)
+
+        def _handle_features(
+            features: tp.Optional[Features], target_id_map: IdMap, dataset_id_map: IdMap
+        ) -> tp.Tuple[tp.Optional[Features], IdMap]:
+            if features is None:
+                return None, target_id_map
+
+            if keep_features_for_removed_entities:
+                all_features_ids = np.arange(len(features))
+                target_id_map = target_id_map.add_ids(all_features_ids, raise_if_already_present=False)
+
+            needed_ids = target_id_map.get_external_sorted_by_internal()
+            features = features.take(needed_ids)
+            return features, target_id_map
+
+        user_features_new, user_id_map = _handle_features(self.user_features, user_id_map, self.user_id_map)
+        item_features_new, item_id_map = _handle_features(self.item_features, item_id_map, self.item_id_map)
+
+        if keep_external_ids:  # external -> 2x internal
+            user_id_map = IdMap(self.user_id_map.convert_to_external(user_id_map.external_ids))
+            item_id_map = IdMap(self.item_id_map.convert_to_external(item_id_map.external_ids))
+
+        filtered_dataset = Dataset(
+            user_id_map=user_id_map,
+            item_id_map=item_id_map,
+            interactions=interactions,
+            user_features=user_features_new,
+            item_features=item_features_new,
+        )
+        return filtered_dataset

rectools/model_selection/cross_validate.py

@@ -14,58 +14,16 @@
 
 import typing as tp
 
-import numpy as np
-import pandas as pd
-
 from rectools.columns import Columns
-from rectools.dataset import Dataset, Features, IdMap, Interactions
+from rectools.dataset import Dataset
 from rectools.metrics import calc_metrics
 from rectools.metrics.base import MetricAtK
-from rectools.models.base import ModelBase
+from rectools.models.base import ErrorBehaviour, ModelBase
 from rectools.types import ExternalIds
 
 from .splitter import Splitter
 
 
-def _gen_2x_internal_ids_dataset(
-    interactions_internal_df: pd.DataFrame,
-    user_features: tp.Optional[Features],
-    item_features: tp.Optional[Features],
-    prefer_warm_inference_over_cold: bool,
-) -> Dataset:
-    """
-    Make new dataset based on given interactions and features from base dataset.
-    Assume that interactions dataframe contains internal ids.
-    Returned dataset contains 2nd level of internal ids.
-    """
-    user_id_map = IdMap.from_values(interactions_internal_df[Columns.User].values)  # 1x internal -> 2x internal
-    item_id_map = IdMap.from_values(interactions_internal_df[Columns.Item].values)  # 1x internal -> 2x internal
-    interactions_train = Interactions.from_raw(interactions_internal_df, user_id_map, item_id_map)  # 2x internal
-
-    def _handle_features(features: tp.Optional[Features], id_map: IdMap) -> tp.Tuple[tp.Optional[Features], IdMap]:
-        if features is None:
-            return None, id_map
-
-        if prefer_warm_inference_over_cold:
-            all_features_ids = np.arange(len(features))  # 1x internal
-            id_map = id_map.add_ids(all_features_ids, raise_if_already_present=False)
-
-        features = features.take(id_map.get_external_sorted_by_internal())  # 2x internal
-        return features, id_map
-
-    user_features_new, user_id_map = _handle_features(user_features, user_id_map)
-    item_features_new, item_id_map = _handle_features(item_features, item_id_map)
-
-    dataset = Dataset(
-        user_id_map=user_id_map,
-        item_id_map=item_id_map,
-        interactions=interactions_train,
-        user_features=user_features_new,
-        item_features=item_features_new,
-    )
-    return dataset
-
-
 def cross_validate(  # pylint: disable=too-many-locals
     dataset: Dataset,
     splitter: Splitter,
@@ -77,6 +35,7 @@ def cross_validate(  # pylint: disable=too-many-locals
     prefer_warm_inference_over_cold: bool = True,
     ref_models: tp.Optional[tp.List[str]] = None,
     validate_ref_models: bool = False,
+    on_unsupported_targets: ErrorBehaviour = "warn",
 ) -> tp.Dict[str, tp.Any]:
     """
     Run cross validation on multiple models with multiple metrics.
@@ -113,6 +72,15 @@ def cross_validate(  # pylint: disable=too-many-locals
     validate_ref_models : bool, default False
         If True include models specified in `ref_models` to all metrics calculations
         and receive their metrics from cross-validation.
+    on_unsupported_targets : Literal["raise", "warn", "ignore"], default "warn"
+        How to handle warm/cold target users when model doesn't support warm/cold inference.
+        Specify "warn" to filter with warning (default in `cross_validate`).
+        Specify "ignore" to filter unsupported targets without a warning.
+        It is highly recommended to pass `CoveredUsers` DQ metric to catch all models with
+        insufficient recommendations for each fold.
+        Specify "raise" to raise ValueError in case unsupported targets are passed. In cross-validation
+        this may cause unexpected errors for some of the complicated models.
+
 
     Returns
     -------
@@ -132,34 +100,26 @@ def cross_validate(  # pylint: disable=too-many-locals
             ]
         }
     """
-    interactions = dataset.interactions
-
-    split_iterator = splitter.split(interactions, collect_fold_stats=True)
+    split_iterator = splitter.split(dataset.interactions, collect_fold_stats=True)
 
     split_infos = []
     metrics_all = []
 
     for train_ids, test_ids, split_info in split_iterator:
         split_infos.append(split_info)
 
-        # ### Prepare split data
-        interactions_df_train = interactions.df.iloc[train_ids]  # 1x internal
-        # We need to avoid fitting models on sparse matrices with all zero rows/columns =>
-        # => we need to create a fold dataset which contains only hot users and items for current training
-        fold_dataset = _gen_2x_internal_ids_dataset(
-            interactions_df_train, dataset.user_features, dataset.item_features, prefer_warm_inference_over_cold
+        fold_dataset = dataset.filter_interactions(
+            row_indexes_to_keep=train_ids,
+            keep_external_ids=True,
+            keep_features_for_removed_entities=prefer_warm_inference_over_cold,
         )
+        interactions_df_test = dataset.interactions.df.loc[test_ids]
+        interactions_df_test[Columns.User] = dataset.user_id_map.convert_to_external(interactions_df_test[Columns.User])
+        interactions_df_test[Columns.Item] = dataset.item_id_map.convert_to_external(interactions_df_test[Columns.Item])
 
-        interactions_df_test = interactions.df.iloc[test_ids]  # 1x internal
-        test_users = interactions_df_test[Columns.User].unique()  # 1x internal
-        catalog = interactions_df_train[Columns.Item].unique()  # 1x internal
-
-        if items_to_recommend is not None:
-            item_ids_to_recommend = dataset.item_id_map.convert_to_internal(
-                items_to_recommend, strict=False
-            )  # 1x internal
-        else:
-            item_ids_to_recommend = None
+        test_users = interactions_df_test[Columns.User].unique()
+        prev_interactions = fold_dataset.get_raw_interactions()
+        catalog = prev_interactions[Columns.Item].unique()
 
         # ### Train ref models if any
         ref_reco = {}
@@ -171,7 +131,8 @@ def cross_validate(  # pylint: disable=too-many-locals
                 dataset=fold_dataset,
                 k=k,
                 filter_viewed=filter_viewed,
-                items_to_recommend=item_ids_to_recommend,
+                items_to_recommend=items_to_recommend,
+                on_unsupported_targets=on_unsupported_targets,
             )
 
         # ### Generate recommendations and calc metrics
@@ -183,19 +144,20 @@ def cross_validate(  # pylint: disable=too-many-locals
                 reco = ref_reco[model_name]
             else:
                 model.fit(fold_dataset)
-                reco = model.recommend(  # 1x internal
+                reco = model.recommend(
                     users=test_users,
                     dataset=fold_dataset,
                     k=k,
                     filter_viewed=filter_viewed,
-                    items_to_recommend=item_ids_to_recommend,
+                    items_to_recommend=items_to_recommend,
+                    on_unsupported_targets=on_unsupported_targets,
                 )
 
             metric_values = calc_metrics(
                 metrics,
                 reco=reco,
                 interactions=interactions_df_test,
-                prev_interactions=interactions_df_train,
+                prev_interactions=prev_interactions,
                 catalog=catalog,
                 ref_reco=ref_reco,
             )