Partition explainer

This notebook demonstrates how to use the Partition explainer, which is an algorithm that uses a hierarchical clustering of the data to recursively partition the input space.

Tabular data example

By default the shap.Explainer interface uses the Parition explainer algorithm only for text and image data, for tabular data the default is to use the Exact or Permutation explainers (depending on how many input features are present). The reason the permutation explainer is preferred over the Partition explainer for tabular data is that the partition explainer algorithm will often leave credit on internal nodes of the hierarchical clustering tree. This works well for text and image data where allocating credit to a contiguous chunk of the input makes sense. But for tabular data this credit sharing between input features in the same cluster can be undesirable. In contrast the Exact and Permutation explainers can still respect a hierarchical clustering tree of the input features while never leaving credit on internal nodes. Note that setting max_evals to infinity will make the Exact explainer, the Permutation explainer, and the Partition explainer all give the same answer, so the tradeoffs here are all about how the SHAP values are approximated with a limited number of samples.

import xgboost
import shap

# train XGBoost model
X,y = shap.datasets.adult()
model = xgboost.XGBClassifier().fit(X, y)

def f(x):
    return model.predict(x, output_margin=True)

# compute SHAP values for the first 500 samples
bg = shap.maskers.Partition(shap.utils.sample(X,100))
explainer = shap.explainers.Partition(f, bg)
shap_values = explainer(X[:500])

explainers.Partition is still in an alpha state, so use with caution...

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
<ipython-input-1-34a8e0adca24> in <module>
     12 bg = shap.maskers.Partition(shap.utils.sample(X,100))
     13 explainer = shap.explainers.Partition(f, bg)
---> 14 shap_values = explainer(X[:500])

~/projects/shap/shap/explainers/_explainer.py in __call__(self, max_evals, main_effects, error_bounds, batch_size, silent, *args, **kwargs)
    188             row_result = self.explain_row(
    189                 *row_args, max_evals=max_evals, main_effects=main_effects, error_bounds=error_bounds,
--> 190                 batch_size=batch_size, silent=silent, **kwargs
    191             )
    192             values.append(row_result.get("values", None))

~/projects/shap/shap/explainers/_partition.py in explain_row(self, max_evals, main_effects, error_bounds, batch_size, silent, *row_args)
    452         output_indexes = None
    453         fixed_context = 1
--> 454         output_indexes, base_value = self.owen(fm, max_evals // 2, output_indexes, fixed_context, batch_size, silent)
    455
    456         if False:

~/projects/shap/shap/explainers/_partition.py in owen(self, fm, npartitions, output_indexes, fixed_context, batch_size, silent)
    497         M = len(fm)
    498         m00 = np.zeros(M, dtype=np.bool)
--> 499         f00 = fm(m00.reshape(1,-1))[0]
    500         base_value = f00
    501         f11 = fm(~m00.reshape(1,-1))[0]

~/projects/shap/shap/utils/_masked_model.py in __call__(self, masks, batch_size)
     51             return self._delta_masking_call(masks, batch_size=batch_size)
     52         else:
---> 53             return self._full_masking_call(masks, batch_size=batch_size)
     54
     55     def _full_masking_call(self, masks, batch_size=None):

~/projects/shap/shap/utils/_masked_model.py in _full_masking_call(self, masks, batch_size)
    119
    120         joined_masked_inputs = self._stack_inputs(all_masked_inputs)
--> 121         outputs = self.model(*joined_masked_inputs)
    122         _assert_output_input_match(joined_masked_inputs, outputs)
    123

~/projects/shap/shap/explainers/_partition.py in <lambda>(x)
     80         self.output_names = output_names
     81
---> 82         self.model = lambda x: np.array(model(x))
     83         self.expected_value = None
     84         if getattr(self.masker, "clustering", None) is None:

<ipython-input-1-34a8e0adca24> in f(x)
      7
      8 def f(x):
----> 9     return model.predict(x, output_margin=True)
     10
     11 # compute SHAP values for the first 500 samples

~/anaconda3/lib/python3.7/site-packages/xgboost/sklearn.py in predict(self, data, output_margin, ntree_limit, validate_features, base_margin)
    885             output_margin=output_margin,
    886             ntree_limit=ntree_limit,
--> 887             validate_features=validate_features)
    888         if output_margin:
    889             # If output_margin is active, simply return the scores

~/anaconda3/lib/python3.7/site-packages/xgboost/core.py in predict(self, data, output_margin, ntree_limit, pred_leaf, pred_contribs, approx_contribs, pred_interactions, validate_features, training)
   1440
   1441         if validate_features:
-> 1442             self._validate_features(data)
   1443
   1444         length = c_bst_ulong()

~/anaconda3/lib/python3.7/site-packages/xgboost/core.py in _validate_features(self, data)
   1852
   1853                 raise ValueError(msg.format(self.feature_names,
-> 1854                                             data.feature_names))
   1855
   1856     def get_split_value_histogram(self, feature, fmap='', bins=None,

ValueError: feature_names mismatch: ['Age', 'Workclass', 'Education-Num', 'Marital Status', 'Occupation', 'Relationship', 'Race', 'Sex', 'Capital Gain', 'Capital Loss', 'Hours per week', 'Country'] ['f0', 'f1', 'f2', 'f3', 'f4', 'f5', 'f6', 'f7', 'f8', 'f9', 'f10', 'f11']
expected Occupation, Education-Num, Capital Gain, Sex, Country, Marital Status, Race, Capital Loss, Age, Workclass, Hours per week, Relationship in input data
training data did not have the following fields: f3, f10, f0, f9, f7, f1, f6, f5, f4, f8, f2, f11

import xgboost
import shap

# train XGBoost model
X,y = shap.datasets.adult()
model = xgboost.XGBClassifier().fit(X, y)

def f(x):
    return model.predict(x, output_margin=True)

# compute SHAP values for the first 500 samples
bg = shap.maskers.Partition(shap.utils.sample(X,100))
explainer = shap.Explainer(f, bg)
shap_values = explainer(X[:500])

shap.plots.bar(shap_values)

shap.plots.bar(shap_values)

shap.plots.heatmap(shap_values)

shap.plots.heatmap(shap_values, instance_order=shap.order.sum)

shap.plots.heatmap(shap_values)

shap.maskers.Par