Scaling Shapley Value computation using Spark
Project description
Shparkley is a PySpark implementation of Shapley values which uses a monte-carlo approximation 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
class MyShparkleyModel(ShparkleyModel):
"""
You need to wrap your model with the ShparkleyModel interface.
"""
def get_required_features(self):
# type: () -> Set[str]
"""
Needs to return a set of feature names for the model.
"""
return ['feature-1', 'feature-2', 'feature-3']
def predict(self, feature_matrix):
# type: (List[Dict[str, Any]]) -> List[float]
"""
Wrapper function to convert the feature matrix into an acceptable format for your model.
This function should return the predicted probabilities.
The feature_matrix is a list of feature dictionaries.
Each dictionary has a mapping from the feature name to the value.
:return: Model predictions for all feature vectors
"""
# Convert the feature matrix into an appropriate form for your model object.
pd_df = pd.DataFrame.from_dict(feature_matrix)
preds = self._model.my_predict(pd_df)
return preds
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 = posv7_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'
)
Project details
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
Source Distribution
shparkley-0.0.2.tar.gz
(5.7 kB
view hashes)
Built Distribution
Close
Hashes for shparkley-0.0.2-py3-none-any.whl
Algorithm | Hash digest | |
---|---|---|
SHA256 | 0d485285ac188ff1fa201ea77cbb237948b41d53afcf2fe1a26111b6f7eeb951 |
|
MD5 | 8bc33931d376957da73c7f32c5195881 |
|
BLAKE2b-256 | 997d6531822f78f76e877122d8d106156bb0b61c17660cf5c26789070639ed7c |