Skip to content

Insertion

Bases: MultitaskMetricMixin, InsertionDeletionBase

The insertion metric measures the faithfulness of an attribution method by evaluating how the prediction score of a model improves when the most important elements of a sequence are gradually added. The importance of the elements is determined by the attribution-based method.

Insertion was introduced by Petsiuk et al. (2017) for image classification. This is our adaptation to the metric for text and in particular for generation (multi-predictions).

A curve is built by computing the prediction score while iteratively inserting the most important elements, starting from a masked sequence. The scores are the softmax outputs, between 0 and 1. The area under this curve (AUC) is then computed to quantify the faithfulness of the attribution method. A higher AUC is better.

The evaluate method returns both:

  • the average AUC across all sequences and targets,
  • for each sequence-target pair, the softmax scores associated to the successive insertions. The softmax scores are preferred over logits as they are bounded between 0 and 1, which makes the AUC more interpretable.

An attribution method is considered good if the Insertion AUC is high, meaning that the model's prediction score increases significantly as the most important elements are added back to the sequence. Conversely, a low AUC indicates that the attribution method is not effective in identifying the most important elements for the model's prediction.

This metric only evaluates the order of importance of the elements in the sequence, not their actual values.

Reference: Vitali Petsiuk, Abir Das, and Kate Saenko. (2017). RISE: Randomized Input Sampling for Explanation of Black-box Models Paper

Examples:

>>> from interpreto.attributions.metrics import Insertion
>>>
>>> # Get explanations from an attribution method
>>> explainer = Method(model, tokenizer, kwargs)
>>> explanations = explainer(inputs, targets)
>>>
>>> # Run the insertion metric
>>> metric = Insertion(model, tokenizer, n_perturbations=100)
>>> auc, metric_scores = metric.evaluate(explanations)

Parameters:

Name Type Description Default

model

PreTrainedModel

model used to generate explanations

 required

tokenizer

PreTrainedTokenizer

Hugging Face tokenizer associated with the model

 required

batch_size

int

batch size for the inference of the metric

 4

granularity

Granularity

granularity level of the perturbations (token, word, sentence, etc.)

 required

device

device

device on which the attribution method will be run

 None

n_perturbations

int

number of perturbations from which the metric will be computed (i.e. the number of steps from which the AUC will be computed).

 100

max_percentage_perturbed

float

maximum percentage of elements in the sequence to be perturbed. Defaults to 1.0, meaning that all elements can be perturbed. If set to 0.5, only half of the elements will be perturbed (i.e. only the 50% most important). This is useful to avoid perturbing too many elements with low scores in long sequences.

 1.0