Add doc for interpretability (#369)

* add doc for interpretability
* update the shap dependency to the release version and point specific version to doc test

Author: heimengqi

azure-pipelines.yml

@@ -64,10 +64,10 @@ jobs:
           displayName: 'Install graphviz'
 
         - script: 'pip install sklearn-contrib-lightning'
-          displayName: 'Install lightning'
-          
-        - script: 'pip install --force-reinstall --no-cache-dir shap'
-          displayName: 'Install public shap'
+          displayName: 'Install lightning'        
+
+        - script: 'pip install git+https://github.com/slundberg/shap.git@d1d2700acc0259f211934373826d5ff71ad514de'
+          displayName: 'Install specific version of shap'  
 
         - script: 'python setup.py build_sphinx -W'
           displayName: 'Build documentation'

doc/reference.rst

@@ -27,6 +27,7 @@ Private Module Reference
     econml._ortho_learner
     econml._cate_estimator
     econml._causal_tree
+    econml._shap
     econml.dml._rlearner
     econml.grf._base_grf
     econml.grf._base_grftree

doc/spec/interpretability.rst

@@ -0,0 +1,99 @@
+Interpretability
+================
+
+Our package offers multiple interpretability tools to better understand the final model CATE.
+
+
+Tree Interpreter
+----------------
+
+Tree Interpreter provides a presentation-ready summary of the key features that explain the biggest differences in responsiveness to an intervention.
+
+:class:`.SingleTreeCateInterpreter` trains a single shallow decision tree for the treatment effect :math:`\theta(X)` you learned from any of
+our available CATE estimators on a small set of feature :math:`X` that you are interested to learn heterogeneity from. The model will split on the cutoff
+points that maximize the treatment effect difference in each leaf. Finally each leaf will be a subgroup of samples that respond to a treatment differently
+from other leaves. 
+
+For instance: 
+
+.. testsetup::
+
+    import numpy as np
+    X = np.random.choice(np.arange(5), size=(100,3))
+    Y = np.random.normal(size=(100,2))
+    y = np.random.normal(size=(100,))
+    T = np.random.choice(np.arange(3), size=(100,2))
+    t = T[:,0]
+    W = np.random.normal(size=(100,2))
+    
+
+.. testcode::
+
+    from econml.cate_interpreter import SingleTreeCateInterpreter
+    from econml.dml import LinearDML
+    est = LinearDML()
+    est.fit(y, t, X=X, W=W)
+    intrp = SingleTreeCateInterpreter(include_model_uncertainty=True, max_depth=2, min_samples_leaf=10)
+    # We interpret the CATE model's behavior based on the features used for heterogeneity
+    intrp.interpret(est, X)
+    # Plot the tree
+    intrp.plot(feature_names=['A', 'B', 'C'], fontsize=12)
+
+Policy Interpreter
+------------------
+Policy Interpreter offers similar functionality but taking cost into consideration. 
+
+Instead of fitting a tree to learn groups that have a different treatment effect, :class:`.SingleTreePolicyInterpreter` tries to split the samples into different treatment groups.
+So in the case of binary treatments it tries to create sub-groups such that all samples within the group have either all positive effect or all negative effect. Thus it tries to
+separate responders from non-responders, as opposed to trying to find groups that have different levels of response.
+
+This way you can construct an interpretable personalized policy where you treat the groups with a postive effect and don't treat the group with a negative effect.
+Our policy tree provides the recommended treatment at each leaf node.
+
+
+For instance: 
+
+.. testcode::
+
+    from econml.cate_interpreter import SingleTreePolicyInterpreter
+    # We find a tree-based treatment policy based on the CATE model
+    # sample_treatment_costs is the cost of treatment. Policy will treat if effect is above this cost.
+    intrp = SingleTreePolicyInterpreter(risk_level=None, max_depth=2, min_samples_leaf=1,min_impurity_decrease=.001)
+    intrp.interpret(est, X, sample_treatment_costs=0.02)
+    # Plot the tree
+    intrp.plot(feature_names=['A', 'B', 'C'], fontsize=12)
+
+
+SHAP
+----
+
+`SHAP <https://shap.readthedocs.io/en/latest/>`_ is a popular open source library for interpreting black-box machine learning
+models using the Shapley values methodology (see e.g. [Lundberg2017]_).
+
+Similar to how black-box predictive machine learning models can be explained with SHAP, we can also explain black-box effect
+heterogeneity models. This approach provides an explanation as to why a heterogeneous causal effect model produced larger or
+smaller effect values for particular segments of the population. Which were the features that lead to such differentiation?
+This question is easy to address when the model is succinctly described, such as the case of linear heterogneity models, 
+where one can simply investigate the coefficients of the model. However, it becomes hard when one starts using more expressive
+models, such as Random Forests and Causal Forests to model effect hetergoeneity. SHAP values can be of immense help to
+understand the leading factors of effect hetergoeneity that the model picked up from the training data.
+
+Our package offers seamless integration with the SHAP library. Every CATE estimator has a method `shap_values`, which returns the
+SHAP value explanation of the estimators output for every treatment and outcome pair. These values can then be visualized with
+the plethora of visualizations that the SHAP library offers. Moreover, whenever possible our library invokes fast specialized
+algorithms from the SHAP library, for each type of final model, which can greatly reduce computation times.
+
+For instance:
+
+.. testcode::
+
+    import shap
+    from econml.dml import LinearDML
+    est = LinearDML()
+    est.fit(y, t, X=X, W=W)
+    shap_values = est.shap_values(X)
+    # local view: explain hetergoeneity for a given observation
+    ind=0
+    shap.plots.force(shap_values["Y0"]["T0"][ind], matplotlib=True)
+    # global view: explain hetergoeneity for a sample of dataset
+    shap.summary_plot(shap_values['Y0']['T0'])

doc/spec/references.rst

@@ -108,4 +108,9 @@ References
 .. [Friedberg2018]
     Friedberg, R., Tibshirani, J., Athey, S., & Wager, S. (2018).
     Local linear forests.
-    arXiv preprint arXiv:1807.11408.
+    arXiv preprint arXiv:1807.11408.
+
+.. [Lundberg2017]
+    Lundberg, S.,  Lee, S. (2017).
+    A Unified Approach to Interpreting Model Predictions.
+    URL https://arxiv.org/abs/1705.07874

doc/spec/spec.rst

@@ -20,6 +20,7 @@ The EconML Python SDK, developed by the ALICE team at MSR New England, incorpora
     estimation
     estimation_iv
     inference
+    interpretability
     references
 
 .. todo::

econml/_shap.py

@@ -1,6 +1,17 @@
 # Copyright (c) Microsoft Corporation. All rights reserved.
 # Licensed under the MIT License.
 
+"""Helper functions to get shap values for different cate estimators.
+
+References
+----------
+Scott Lundberg, Su-In Lee (2017)
+    A Unified Approach to Interpreting Model Predictions.
+    NeurIPS, https://arxiv.org/abs/1705.07874
+
+
+"""
+
 import shap
 from collections import defaultdict
 import numpy as np

setup.cfg

@@ -50,7 +50,7 @@ install_requires =
     graphviz
     matplotlib
     pandas < 1.1
-    shap @ git+https://github.com/slundberg/shap.git
+    shap ~= 0.38.1
 test_suite = econml.tests
 tests_require =
     pytest