Vasilis/casestudy

azure-pipelines-steps.yml

@@ -54,7 +54,7 @@ jobs:
     condition: and(succeeded(), eq(variables['Agent.OS'], 'Linux'))
 
   # Install the package
-  - script: 'python -m pip install --upgrade pip && pip install --upgrade setuptools wheel Cython && pip install ${{ parameters.package }}'
+  - script: 'pip install --upgrade setuptools wheel Cython && pip install ${{ parameters.package }}'
     displayName: 'Install dependencies'
 
   - ${{ parameters.job.steps }}

econml/dml/causal_forest.py

@@ -73,7 +73,7 @@ def fit(self, X, T, T_res, Y_res, sample_weight=None, freq_weight=None, sample_v
                      "where available.")
             residuals = Y_res - np.einsum('ijk,ik->ij', oob_preds, T_res)
             propensities = T - T_res
-            VarT = propensities * (1 - propensities)
+            VarT = np.clip(propensities * (1 - propensities), 1e-10, np.inf)
             drpreds = oob_preds
             drpreds += cross_product(residuals, T_res / VarT).reshape((-1, Y_res.shape[1], T_res.shape[1]))
             drpreds[np.isnan(oob_preds)] = np.nan

econml/policy/_drlearner.py

@@ -4,7 +4,7 @@
 from warnings import warn
 import numpy as np
 from sklearn.base import clone
-from ..utilities import check_inputs, filter_none_kwargs
+from ..utilities import check_inputs, filter_none_kwargs, check_input_arrays
 from ..dr import DRLearner
 from ..dr._drlearner import _ModelFinal
 from .._tree_exporter import _SingleTreeExporterMixin
@@ -98,6 +98,24 @@ def predict_value(self, X):
         """
         return self.drlearner_.const_marginal_effect(X)
 
+    def predict_proba(self, X):
+        """ Predict the probability of recommending each treatment
+
+        Parameters
+        ----------
+        X : array-like of shape (n_samples, n_features)
+            The input samples.
+
+        Returns
+        -------
+        treatment_proba : array-like of shape (n_samples, n_treatments)
+            The probability of each treatment recommendation
+        """
+        X, = check_input_arrays(X)
+        if self.drlearner_.featurizer_ is not None:
+            X = self.drlearner_.featurizer_.fit_transform(X)
+        return self.policy_model_.predict_proba(X)
+
     def predict(self, X):
         """ Get recommended treatment for each sample.
 
@@ -111,9 +129,11 @@ def predict(self, X):
         treatment : array-like of shape (n_samples,)
             The index of the recommended treatment in the same order as in categories, or in
             lexicographic order if `categories='auto'`. 0 corresponds to the baseline/control treatment.
+            For ensemble policy models, recommended treatments are aggregated from each model in the ensemble
+            and the treatment that receives the most votes is returned. Use `predict_proba` to get the fraction
+            of models in the ensemble that recommend each treatment for each sample.
         """
-        values = self.predict_value(X)
-        return np.argmax(np.hstack([np.zeros((values.shape[0], 1)), values]), axis=1)
+        return np.argmax(self.predict_proba(X), axis=1)
 
     def policy_feature_names(self, *, feature_names=None):
         """

econml/policy/_forest/_forest.py

@@ -455,6 +455,43 @@ def predict_value(self, X):
 
         return y_hat
 
+    def predict_proba(self, X):
+        """ Predict the probability of recommending each treatment
+
+        Parameters
+        ----------
+        X : {array-like} of shape (n_samples, n_features)
+            The input samples. Internally, it will be converted to
+            ``dtype=np.float64``.
+        check_input : bool, default=True
+            Allow to bypass several input checking.
+            Don't use this parameter unless you know what you do.
+
+        Returns
+        -------
+        treatment_proba : array-like of shape (n_samples, n_treatments)
+            The probability of each treatment recommendation
+        """
+        check_is_fitted(self)
+        # Check data
+        X = self._validate_X_predict(X)
+
+        # Assign chunk of trees to jobs
+        n_jobs, _, _ = _partition_estimators(self.n_estimators, self.n_jobs)
+
+        # avoid storing the output of every estimator by summing them here
+        y_hat = np.zeros((X.shape[0], self.n_outputs_), dtype=np.float64)
+
+        # Parallel loop
+        lock = threading.Lock()
+        Parallel(n_jobs=n_jobs, verbose=self.verbose, require="sharedmem")(
+            delayed(_accumulate_prediction)(e.predict_proba, X, [y_hat], lock)
+            for e in self.estimators_)
+
+        y_hat /= len(self.estimators_)
+
+        return y_hat
+
     def predict(self, X):
         """ Predict the best treatment for each sample
 
@@ -467,6 +504,9 @@ def predict(self, X):
         Returns
         -------
         treatment : array-like of shape (n_samples)
-            The recommded treatment, i.e. the treatment index with the largest reward for each sample
+            The recommded treatment, i.e. the treatment index most often predicted to have the highest reward
+            for each sample. Recommended treatments are aggregated from each tree in the ensemble and the treatment
+            that receives the most votes is returned. Use `predict_proba` to get the fraction of trees in the ensemble
+            that recommend each treatment for each sample.
         """
-        return np.argmax(self.predict_value(X), axis=1)
+        return np.argmax(self.predict_proba(X), axis=1)

econml/policy/_forest/_tree.py

@@ -261,6 +261,30 @@ def predict(self, X, check_input=True):
         pred = self.tree_.predict(X)
         return np.argmax(pred, axis=1)
 
+    def predict_proba(self, X, check_input=True):
+        """ Predict the probability of recommending each treatment
+
+        Parameters
+        ----------
+        X : {array-like} of shape (n_samples, n_features)
+            The input samples. Internally, it will be converted to
+            ``dtype=np.float64``.
+        check_input : bool, default=True
+            Allow to bypass several input checking.
+            Don't use this parameter unless you know what you do.
+
+        Returns
+        -------
+        treatment_proba : array-like of shape (n_samples, n_treatments)
+            The probability of each treatment recommendation
+        """
+        check_is_fitted(self)
+        X = self._validate_X_predict(X, check_input)
+        pred = self.tree_.predict(X)
+        proba = np.zeros(pred.shape)
+        proba[np.arange(X.shape[0]), np.argmax(pred, axis=1)] = 1
+        return proba
+
     def predict_value(self, X, check_input=True):
         """ Predict the expected value of each treatment for each sample
 

econml/tests/test_policy_forest.py

@@ -290,6 +290,8 @@ def test_non_standard_input(self,):
         forest = PolicyForest(n_estimators=20, n_jobs=1, random_state=123).fit(X, y)
         pred = forest.predict(X)
         pred_val = forest.predict_value(X)
+        pred_prob = forest.predict_proba(X)
+        assert pred_prob.shape == (X.shape[0], 2)
         feat_imp = forest.feature_importances()
         forest = PolicyForest(n_estimators=20, n_jobs=1, random_state=123).fit(X.astype(np.float32),
                                                                                np.asfortranarray(y))
@@ -298,12 +300,15 @@ def test_non_standard_input(self,):
         forest = PolicyForest(n_estimators=20, n_jobs=1, random_state=123).fit(tuple(X), tuple(y))
         np.testing.assert_allclose(pred, forest.predict(tuple(X)))
         np.testing.assert_allclose(pred_val, forest.predict_value(tuple(X)))
+        np.testing.assert_allclose(pred_prob, forest.predict_proba(tuple(X)))
         forest = PolicyForest(n_estimators=20, n_jobs=1, random_state=123).fit(list(X), list(y))
         np.testing.assert_allclose(pred, forest.predict(list(X)))
         np.testing.assert_allclose(pred_val, forest.predict_value(list(X)))
+        np.testing.assert_allclose(pred_prob, forest.predict_proba(list(X)))
         forest = PolicyForest(n_estimators=20, n_jobs=1, random_state=123).fit(pd.DataFrame(X), pd.DataFrame(y))
         np.testing.assert_allclose(pred, forest.predict(pd.DataFrame(X)))
         np.testing.assert_allclose(pred_val, forest.predict_value(pd.DataFrame(X)))
+        np.testing.assert_allclose(pred_prob, forest.predict_proba(pd.DataFrame(X)))
 
         groups = np.repeat(np.arange(X.shape[0]), 2)
         Xraw = X.copy()
@@ -324,6 +329,15 @@ def test_non_standard_input(self,):
                                    forest.predict_value(Xraw[mask]).flatten(), atol=.08)
         np.testing.assert_allclose(feat_imp, forest.feature_importances(), atol=1e-4)
         np.testing.assert_allclose(feat_imp, forest.feature_importances_, atol=1e-4)
+        pred = forest.predict(X)
+        pred_val = forest.predict_value(X)
+        pred_prob = forest.predict_proba(X)
+        np.testing.assert_allclose(pred, forest.predict(tuple(X)))
+        np.testing.assert_allclose(pred_val, forest.predict_value(tuple(X)))
+        np.testing.assert_allclose(pred, forest.predict(pd.DataFrame(X)))
+        np.testing.assert_allclose(pred_val, forest.predict_value(pd.DataFrame(X)))
+        np.testing.assert_allclose(pred_prob, forest.predict_proba(pd.DataFrame(X)))
+
         return
 
     def test_raise_exceptions(self,):

setup.cfg

@@ -53,6 +53,7 @@ tests_require =
     seaborn
     lightgbm
     xgboost
+    jupyter-client <= 6.1.12 
 
 [options.extras_require]
 automl =