Implement forecast decomposition for linear models

CHANGELOG.md

@@ -7,6 +7,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## Unreleased
 ### Added
+- Method `predict_components` for forecast decomposition in `_SklearnAdapter` and `_LinearAdapter` for linear models ([#1164](https://github.com/tinkoff-ai/etna/pull/1164))
 - Target components logic into base classes of models ([#1158](https://github.com/tinkoff-ai/etna/pull/1158))
 - Target components logic to TSDataset ([#1153](https://github.com/tinkoff-ai/etna/pull/1153))
 - Methods `save` and `load` to HierarchicalPipeline ([#1096](https://github.com/tinkoff-ai/etna/pull/1096))

etna/models/linear.py

@@ -1,12 +1,52 @@
+import pandas as pd
 from sklearn.linear_model import ElasticNet
 from sklearn.linear_model import LinearRegression
 
-from etna.models.sklearn import SklearnMultiSegmentModel
-from etna.models.sklearn import SklearnPerSegmentModel
+from etna.models.base import NonPredictionIntervalContextIgnorantAbstractModel
+from etna.models.mixins import MultiSegmentModelMixin
+from etna.models.mixins import NonPredictionIntervalContextIgnorantModelMixin
+from etna.models.mixins import PerSegmentModelMixin
+from etna.models.sklearn import _SklearnAdapter
 
 
-class LinearPerSegmentModel(SklearnPerSegmentModel):
-    """Class holding per segment :py:class:`sklearn.linear_model.LinearRegression`."""
+class _LinearAdapter(_SklearnAdapter):
+    def predict_components(self, df: pd.DataFrame) -> pd.DataFrame:
+        """Estimate prediction components.
+
+        Parameters
+        ----------
+        df:
+            features dataframe
+
+        Returns
+        -------
+        :
+            dataframe with prediction components
+        """
+        if self.regressor_columns is None:
+            raise ValueError("Model is not fitted! Fit the model before estimating forecast components!")
+
+        components_coefs = self.model.coef_
+        target_components = df[self.model.feature_names_in_].apply(pd.to_numeric)
+        target_components = components_coefs * target_components
+        if self.model.fit_intercept:
+            target_components["intercept"] = self.model.intercept_
+        target_components = target_components.add_prefix("target_component_")
+        return target_components
+
+
+class LinearPerSegmentModel(
+    PerSegmentModelMixin,
+    NonPredictionIntervalContextIgnorantModelMixin,
+    NonPredictionIntervalContextIgnorantAbstractModel,
+):
+    """
+    Class holding per segment :py:class:`sklearn.linear_model.LinearRegression`.
+
+    Notes
+    -----
+    Target components are formed as the terms from linear regression formula.
+    """
 
     def __init__(self, fit_intercept: bool = True, **kwargs):
         """
@@ -20,11 +60,23 @@ def __init__(self, fit_intercept: bool = True, **kwargs):
         """
         self.fit_intercept = fit_intercept
         self.kwargs = kwargs
-        super().__init__(regressor=LinearRegression(fit_intercept=self.fit_intercept, **self.kwargs))
+        super().__init__(
+            base_model=_LinearAdapter(regressor=LinearRegression(fit_intercept=self.fit_intercept, **self.kwargs))
+        )
+
 
+class ElasticPerSegmentModel(
+    PerSegmentModelMixin,
+    NonPredictionIntervalContextIgnorantModelMixin,
+    NonPredictionIntervalContextIgnorantAbstractModel,
+):
+    """
+    Class holding per segment :py:class:`sklearn.linear_model.ElasticNet`.
 
-class ElasticPerSegmentModel(SklearnPerSegmentModel):
-    """Class holding per segment :py:class:`sklearn.linear_model.ElasticNet`."""
+    Notes
+    -----
+    Target components are formed as the terms from linear regression formula.
+    """
 
     def __init__(self, alpha: float = 1.0, l1_ratio: float = 0.5, fit_intercept: bool = True, **kwargs):
         """
@@ -55,17 +107,29 @@ def __init__(self, alpha: float = 1.0, l1_ratio: float = 0.5, fit_intercept: boo
         self.fit_intercept = fit_intercept
         self.kwargs = kwargs
         super().__init__(
-            regressor=ElasticNet(
-                alpha=self.alpha,
-                l1_ratio=self.l1_ratio,
-                fit_intercept=self.fit_intercept,
-                **self.kwargs,
+            base_model=_LinearAdapter(
+                regressor=ElasticNet(
+                    alpha=self.alpha,
+                    l1_ratio=self.l1_ratio,
+                    fit_intercept=self.fit_intercept,
+                    **self.kwargs,
+                )
             )
         )
 
 
-class LinearMultiSegmentModel(SklearnMultiSegmentModel):
-    """Class holding :py:class:`sklearn.linear_model.LinearRegression` for all segments."""
+class LinearMultiSegmentModel(
+    MultiSegmentModelMixin,
+    NonPredictionIntervalContextIgnorantModelMixin,
+    NonPredictionIntervalContextIgnorantAbstractModel,
+):
+    """
+    Class holding :py:class:`sklearn.linear_model.LinearRegression` for all segments.
+
+    Notes
+    -----
+    Target components are formed as the terms from linear regression formula.
+    """
 
     def __init__(self, fit_intercept: bool = True, **kwargs):
         """
@@ -79,11 +143,23 @@ def __init__(self, fit_intercept: bool = True, **kwargs):
         """
         self.fit_intercept = fit_intercept
         self.kwargs = kwargs
-        super().__init__(regressor=LinearRegression(fit_intercept=self.fit_intercept, **self.kwargs))
+        super().__init__(
+            base_model=_LinearAdapter(regressor=LinearRegression(fit_intercept=self.fit_intercept, **self.kwargs))
+        )
+
 
+class ElasticMultiSegmentModel(
+    MultiSegmentModelMixin,
+    NonPredictionIntervalContextIgnorantModelMixin,
+    NonPredictionIntervalContextIgnorantAbstractModel,
+):
+    """
+    Class holding :py:class:`sklearn.linear_model.ElasticNet` for all segments.
 
-class ElasticMultiSegmentModel(SklearnMultiSegmentModel):
-    """Class holding :py:class:`sklearn.linear_model.ElasticNet` for all segments."""
+    Notes
+    -----
+    Target components are formed as the terms from linear regression formula.
+    """
 
     def __init__(self, alpha: float = 1.0, l1_ratio: float = 0.5, fit_intercept: bool = True, **kwargs):
         """
@@ -114,10 +190,12 @@ def __init__(self, alpha: float = 1.0, l1_ratio: float = 0.5, fit_intercept: boo
         self.fit_intercept = fit_intercept
         self.kwargs = kwargs
         super().__init__(
-            regressor=ElasticNet(
-                alpha=self.alpha,
-                l1_ratio=self.l1_ratio,
-                fit_intercept=self.fit_intercept,
-                **self.kwargs,
+            base_model=_LinearAdapter(
+                regressor=ElasticNet(
+                    alpha=self.alpha,
+                    l1_ratio=self.l1_ratio,
+                    fit_intercept=self.fit_intercept,
+                    **self.kwargs,
+                )
             )
         )

etna/models/sklearn.py

@@ -63,6 +63,21 @@ def predict(self, df: pd.DataFrame) -> np.ndarray:
         pred = self.model.predict(features)
         return pred
 
+    def predict_components(self, df: pd.DataFrame) -> pd.DataFrame:
+        """Estimate prediction components.
+
+        Parameters
+        ----------
+        df:
+            features dataframe
+
+        Returns
+        -------
+        :
+            dataframe with prediction components
+        """
+        raise NotImplementedError("Prediction decomposition isn't currently implemented!")
+
     def get_model(self) -> RegressorMixin:
         """Get internal sklearn model that is used inside etna class.
 

tests/test_models/test_linear_model.py

@@ -1,3 +1,6 @@
+from typing import List
+from typing import Tuple
+
 import numpy as np
 import pandas as pd
 import pytest
@@ -9,6 +12,7 @@
 from etna.models.linear import ElasticPerSegmentModel
 from etna.models.linear import LinearMultiSegmentModel
 from etna.models.linear import LinearPerSegmentModel
+from etna.models.linear import _LinearAdapter
 from etna.pipeline import Pipeline
 from etna.transforms.math import LagTransform
 from etna.transforms.timestamp import DateFlagsTransform
@@ -42,6 +46,14 @@ def ts_with_categoricals(random_seed) -> TSDataset:
     return ts
 
 
+@pytest.fixture
+def df_with_regressors(example_tsds) -> Tuple[pd.DataFrame, List[str]]:
+    lags = LagTransform(in_column="target", lags=[7], out_column="lag")
+    dflg = DateFlagsTransform(day_number_in_week=True, day_number_in_month=True, is_weekend=False, out_column="df")
+    example_tsds.fit_transform([lags, dflg])
+    return example_tsds.to_pandas(flatten=True).dropna(), example_tsds.regressors
+
+
 def linear_segments_by_parameters(alpha_values, intercept_values):
     dates = pd.date_range(start="2020-02-01", freq="D", periods=210)
     x = np.arange(210)
@@ -269,3 +281,47 @@ def test_save_load(model, example_tsds):
     horizon = 3
     transforms = [LagTransform(in_column="target", lags=list(range(horizon, horizon + 3)))]
     assert_model_equals_loaded_original(model=model, ts=example_tsds, transforms=transforms, horizon=horizon)
+
+
+@pytest.mark.parametrize(
+    "fit_intercept, expected_component_names",
+    [
+        (
+            True,
+            [
+                "target_component_lag_7",
+                "target_component_df_day_number_in_week",
+                "target_component_df_day_number_in_month",
+                "target_component_intercept",
+            ],
+        ),
+        (
+            False,
+            [
+                "target_component_lag_7",
+                "target_component_df_day_number_in_week",
+                "target_component_df_day_number_in_month",
+            ],
+        ),
+    ],
+)
+@pytest.mark.parametrize("regressor_constructor", (LinearRegression, ElasticNet))
+def test_linear_adapter_predict_components_correct_names(
+    df_with_regressors, regressor_constructor, fit_intercept, expected_component_names
+):
+    df, regressors = df_with_regressors
+    adapter = _LinearAdapter(regressor=regressor_constructor(fit_intercept=fit_intercept))
+    adapter.fit(df=df, regressors=regressors)
+    target_components = adapter.predict_components(df)
+    assert sorted(target_components.columns) == sorted(expected_component_names)
+
+
+@pytest.mark.parametrize("fit_intercept", (True, False))
+@pytest.mark.parametrize("regressor_constructor", (LinearRegression, ElasticNet))
+def test_linear_adapter_predict_components_sum_up_to_target(df_with_regressors, regressor_constructor, fit_intercept):
+    df, regressors = df_with_regressors
+    adapter = _LinearAdapter(regressor=regressor_constructor(fit_intercept=fit_intercept))
+    adapter.fit(df=df, regressors=regressors)
+    target = adapter.predict(df)
+    target_components = adapter.predict_components(df)
+    np.testing.assert_array_almost_equal(target, target_components.sum(axis=1), decimal=10)