Sobol Attribution

Bases: MultitaskExplainerMixin, AttributionExplainer

Sobol is a variance-based sensitivity analysis method used to quantify the contribution of each input component to the output variance of the model.

It estimates both the first-order (main) and total (interaction) effects of features using Monte Carlo sampling strategies. In NLP, Sobol helps assess which words or tokens are most influential for the model’s decision, including how they interact with one another.

Reference: Fel et al. (2021). Look at the variance! Efficient black-box explanations with Sobol-based sensitivity analysis. Paper

Examples:

>>> from interpreto import Granularity, Sobol
>>> from interpreto.attributions import InferenceModes
>>> method = Sobol(model, tokenizer, batch_size=4,
>>>                inference_mode=InferenceModes.LOGITS,
>>>                n_token_perturbations=8,
>>>                granularity=Granularity.WORD,
>>>                sobol_indices_order=Sobol.sobol_indices_orders.FIRST_ORDER,
>>>                sampler=Sobol.samplers.SOBOL))
>>> 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 (e.g., token, word, sentence).	WORD
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
n_token_perturbations	int	the number of perturbations to generate	32
sobol_indices	SobolIndicesOrders	Sobol indices order, either FIRST_ORDER or TOTAL_ORDER.	required
sampler	SequenceSamplers	Sobol sequence sampler, either SOBOL, HALTON or LatinHypercube.	SOBOL
device	device	device on which the attribution method will be run	None