Skip to content

Regression explanations with Xplique

Attributions: Regression and Tabular data tutorial Open In Colab

Which kind of tasks are supported by Xplique?

With the operator's api you can treat many different problems with Xplique. There is one operator for each task.

Task and Documentation link operator parameter value
from xplique.Tasks Enum		 Tutorial link
Classification CLASSIFICATION Open In Colab
Object Detection OBJECT_DETECTION Open In Colab
Regression REGRESSION Open In Colab
Semantic Segmentation SEMANTIC_SEGMENTATION Open In Colab

Info

They all share the API for Xplique attribution methods.

Warning

In Xplique, for now with regression, predictions can only be explained output by output. Indeed, explaining several output simultaneously brings new problematic and we are currently working on an operator to solve this.

Simple example

import xplique
from xplique.attributions import Saliency
from xplique.metrics import Deletion

# load inputs and model
# ...

# for regression, `targets` indicates the output of interest, here output 3
targets = tf.one_hot([2], depth=nb_outputs, axis=-1)

# compute explanations by specifying the regression operator
explainer = Saliency(model, operator=xplique.Tasks.REGRESSION)
explanations = explainer(inputs, targets)

# compute metrics on these explanations
metric = Deletion(model, inputs, targets, operator=xplique.Tasks.REGRESSION)
score_saliency = metric(explanations)

How to use it?

To apply attribution methods, the common API documentation describes the parameters and how to fix them. However, depending on the task and thus on the operator, there are three points that vary:

  • The operator parameter value, it is an Enum or a string identifying the task,

  • The model's output specification, as model(inputs) is used in the computation of the operators, and

  • The targets parameter format, indeed, the targets parameter specifies what to explain and the format of such specification depends on the task.

The operator

How to specify it

In Xplique, to adapt attribution methods, you should specify the task to the operator parameter. In the case of regression, with either: 

Method(model, operator="regression")
# or
Method(model, operator=xplique.Tasks.REGRESSION)

The computation

The regression operator works similarly to the classification operator, it asks for the output of interest via targets and returns this output. See targets section for more detail. 

scores = tf.reduce_sum(model(inputs) * targets, axis=-1)

The behavior

  • In the case of perturbation-based methods, the perturbation score corresponds to the difference between the initial value of the output of interest and the same output for predictions over perturbed inputs.
  • For gradient-based methods, the gradient of the model's predictions for the output of interest.

Model's output

We expect model(inputs) to yield a \((n, d)\) tensor or array where \(n\) is the number of input samples and \(d\) is the number of variables the model should predict (possibly one).

The targets parameter

Role

The targets parameter specifies what is to explain in the inputs, it is passed to the explain or to the __call__ method of an explainer or metric and used by the operators. In the case of regression it indicates which of the output should be explained.

Format

The targets parameter in the case of regression should have the same shape as the model's output as they are multiplied. Hence, the shape is \((n, d)\) with \(n\) the number of samples to be explained (it should match the first dimension of inputs) and \(d\) is the number of variables (possibly one).

In practice

In the simple example, the targets value provided is computed with tf.one_hot. Indeed, the regression operator takes as targets the one hot encoding of the index of the output to explain.