Gradient Shap

Bases: MultitaskExplainerMixin, AttributionExplainer

GradientSHAP is a gradient-based Shapley value estimator that computes attributions by integrating model gradients along a path between a baseline (reference) and the input. It approximates Shapley values by averaging multiple stochastic integrated gradients across randomly sampled paths.

By combining ideas from Integrated Gradients and Shapley value theory, GradientSHAP provides additive feature attributions with strong consistency guarantees, while capturing non-linear effects.

Reference: Lundberg and Lee (2017). A Unified Approach to Interpreting Model Predictions. Paper

Examples:

>>> from interpreto import GradientShap
>>> method = GradientShap(model, tokenizer, batch_size=4,
>>>                       n_perturbations=20,
>>>                       baseline=0,
>>>                       noise_std=0.1,)
>>> explanations = method(text)

Parameters:

Name	Type	Description	Default
model	PreTrainedModel	model to explain	required
tokenizer	PreTrainedTokenizer	Hugging Face tokenizer associated with the model	required
batch_size	int	batch size for the attribution method	4
granularity	Granularity	The level of granularity for the explanation. Options are: ALL_TOKENS, TOKEN, WORD, or SENTENCE. Defaults to Granularity.WORD. To obtain it, from interpreto import Granularity then Granularity.WORD.	WORD
granularity_aggregation_strategy	GranularityAggregationStrategy	how to aggregate token-level attributions into granularity scores. Options are: MEAN, MAX, MIN, SUM, and SIGNED_MAX. Ignored for granularity set to ALL_TOKENS or TOKEN.	MEAN
device	device	device on which the attribution method will be run	None
inference_mode	Callable[[Tensor], Tensor]	The mode used for inference. It can be either one of LOGITS, SOFTMAX, or LOG_SOFTMAX. Use InferenceModes to choose the appropriate mode.	LOGITS
input_x_gradient	bool	If True, multiplies the input embeddings with their gradients before aggregation. Defaults to True.	True
n_perturbations	int	the number of interpolations to generate	10
baseline	Tensor | float | None	the baseline to use for the interpolations	None
noise_std	float	the standard deviation of the noise added to the baseline	0.1