Shparkley: Scaling Shapley Values with Spark

.. inclusion-marker-start-do-not-remove

.. contents::

Shparkley is a PySpark implementation of Shapley values <https://papers.nips.cc/paper/7062-a-unified-approach-to-interpreting-model-predictions.pdf>_ which uses a monte-carlo approximation <https://link.springer.com/article/10.1007/s10115-013-0679-x>_ algorithm.

Given a dataset and machine learning model, Shparkley can compute Shapley values for all features for a feature vector. Shparkley also handles training weights and is model-agnostic.

Installation

pip install shparkley

Requirements

You must have Apache Spark installed on your machine/cluster.

Example Usage

.. code-block:: python

from typing import List

from sklearn.base import ClassifierMixin

from affirm.model_interpretation.shparkley.estimator_interface import OrderedSet, ShparkleyModel
from affirm.model_interpretation.shparkley.spark_shapley import compute_shapley_for_sample


class MyShparkleyModel(ShparkleyModel):
    """
    You need to wrap your model with this interface (by subclassing ShparkleyModel)
    """
    def __init__(self, model: ClassifierMixin, required_features: OrderedSet):
        self._model = model
        self._required_features = required_features

    def predict(self, feature_matrix: List[OrderedDict]) -> List[float]:
        """
        Generates one prediction per row, taking in a list of ordered dictionaries (one per row).
        """
        pd_df = pd.DataFrame.from_dict(feature_matrix)
        preds = self._model.predict_proba(pd_df)[:, 1]
        return preds

    def _get_required_features(self) -> OrderedSet:
        """
        An ordered set of feature column names
        """
        return self._required_features

row = dataset.filter(dataset.row_id == 'xxxx').rdd.first()
shparkley_wrapped_model = MyShparkleyModel(my_model)

# You need to sample your dataset based on convergence criteria.
# More samples results in more accurate shapley values.
# Repartitioning and caching the sampled dataframe will speed up computation.
sampled_df = training_df.sample(0.1, True).repartition(75).cache()

shapley_scores_by_feature = compute_shapley_for_sample(
    df=sampled_df,
    model=shparkley_wrapped_model,
    row_to_investigate=row,
    weight_col_name='training_weight_column_name'
)