SHAP: SHapley Additive exPlanations

The SHAP package is built on the concept of a Shapley value and can generate explanations model-agnostically. So it only requires input and output values, not model internals:

SHAP (SHapley Additive exPlanations) is a game theoretic approach to explain the output of any machine learning model. It connects optimal credit allocation with local explanations using the classic Shapley values from game theory and their related extensions. (README)

Additionally, this package also contains several model-specific implementations of Shapley values that are optimized for a particular machine learning model and sometimes even for a particular library.

If we inspect we see the complete overview of how the explainers can easily be imported (the README does not list all of them at the moment of writing):

# explainers
from .explainers._explainer import Explainer
from .explainers._kernel import Kernel as KernelExplainer
from .explainers._sampling import Sampling as SamplingExplainer
from .explainers._tree import Tree as TreeExplainer
from .explainers._gpu_tree import GPUTree as GPUTreeExplainer
from .explainers._deep import Deep as DeepExplainer
from .explainers._gradient import Gradient as GradientExplainer
from .explainers._linear import Linear as LinearExplainer
from .explainers._partition import Partition as PartitionExplainer
from .explainers._permutation import Permutation as PermutationExplainer
from .explainers._additive import Additive as AdditiveExplainer

Model-specific optimizations

Ensembles of trees

There is specific support for tree (ensemble) models from XGBoost, LightGBM, CatBoost, scikit-learn, PySpark. These models can be passed directly into the shap.Explainer. This particular implementation can compute Shapley values exactly.

The shap.TreeExplainer can also compute SHAP interaction values for pairwise interactions between features, as such: shap.TreeExplainer(model).shap_interaction_values(X).

shap.GPUTreeExplainer is a TreeExplainer optimized for GPU.

Natural Language models

You can generate Shapley values for a NLP transformer pipeline (e.g. one from Hugging Face) by passing the whole pipeline into the default shap.Explainer.

Deep Neural Networks

A shap.DeepExplainer is provided specifically for deep learning models and is an extension of DeepLIFT (see Shrikumar et al., 2018).

  • Support for TensorFlow/Keras and PyTorch (preliminary support at the moment of writing).
  • “DeepLIFT compares the activation of each neuron to its ‘reference activation’ and assigns contribution scores according to the difference. By optionally giving separate consideration to positive and negative contributions, DeepLIFT can also reveal dependencies which are missed by other approaches.” (Shrikumar et al., 2017)

Gradient-based explanations


Gradient-based methods can also be used to explain the effect of (intermediate) deep learning layers on the network’s output. For example for an input image, you can visualize which pixels in a feature map increase the probability of a particular class and which ones decrease that probability.

Expected gradients combines ideas from Integrated Gradients, SHAP, and SmoothGrad into a single expected value equation. This allows an entire dataset to be used as the background distribution (as opposed to a single reference value) and allows local smoothing. If we approximate the model with a linear function between each background data sample and the current input to be explained, and we assume the input features are independent then expected gradients will compute approximate SHAP values. (README)

Model-agnostic explanations


Kernel SHAP uses a specially-weighted local linear regression to estimate SHAP values for any model. (README)

The Shapley values indicate a positive or negative push away from the average model output over the whole training dataset.