explain function

Explain model predictions

Once an explainer has been created using the lime() function it can be used to explain the result of the model on new observations. The explain()

function takes new observation along with the explainer and returns a data.frame with prediction explanations, one observation per row. The returned explanations can then be visualised in a number of ways, e.g. with plot_features().

## S3 method for class 'data.frame'
explain(
  x,
  explainer,
  labels = NULL,
  n_labels = NULL,
  n_features,
  n_permutations = 5000,
  feature_select = "auto",
  dist_fun = "gower",
  kernel_width = NULL,
  gower_pow = 1,
  ...
)

## S3 method for class 'character'
explain(
  x,
  explainer,
  labels = NULL,
  n_labels = NULL,
  n_features,
  n_permutations = 5000,
  feature_select = "auto",
  single_explanation = FALSE,
  ...
)

explain(
  x,
  explainer,
  labels,
  n_labels = NULL,
  n_features,
  n_permutations = 5000,
  feature_select = "auto",
  ...
)

## S3 method for class 'imagefile'
explain(
  x,
  explainer,
  labels = NULL,
  n_labels = NULL,
  n_features,
  n_permutations = 1000,
  feature_select = "auto",
  n_superpixels = 50,
  weight = 20,
  n_iter = 10,
  p_remove = 0.5,
  batch_size = 10,
  background = "grey",
  ...
)

Arguments

• x: New observations to explain, of the same format as used when creating the explainer

• explainer: An explainer object to use for explaining the observations

• labels: The specific labels (classes) to explain in case the model is a classifier. For classifiers either this or n_labels must be given.

• n_labels: The number of labels to explain. If this is given for classifiers the top n_label classes will be explained.

• n_features: The number of features to use for each explanation.

• n_permutations: The number of permutations to use for each explanation.

• feature_select: The algorithm to use for selecting features. One of:

  • "auto": If n_features \<= 6 use "forward_selection" else use "highest_weights".
  • "none": Ignore n_features and use all features.
  • "forward_selection": Add one feature at a time until n_features is reached, based on quality of a ridge regression model.
  • "highest_weights": Fit a ridge regression and select the n_features with the highest absolute weight.
  • "lasso_path": Fit a lasso model and choose the n_features whose lars path converge to zero the latest.
  • "tree" : Fit a tree to select n_features (which needs to be a power of 2). It requires last version of XGBoost.

• dist_fun: The distance function to use for calculating the distance from the observation to the permutations. If dist_fun = 'gower' (default) it will use gower::gower_dist(). Otherwise it will be forwarded to stats::dist()

• kernel_width: The width of the exponential kernel that will be used to convert the distance to a similarity in case dist_fun != 'gower'.

• gower_pow: A modifier for gower distance. The calculated distance will be raised to the power of this value.

• ...: Parameters passed on to the predict_model() method

• single_explanation: A boolean indicating whether to pool all text in x into a single explanation.

• n_superpixels: The number of segments an image should be split into

• weight: How high should locality be weighted compared to colour. High values leads to more compact superpixels, while low values follow the image structure more

• n_iter: How many iterations should the segmentation run for

• p_remove: The probability that a superpixel will be removed in each permutation

• batch_size: The number of explanations to handle at a time

• background: The colour to use for blocked out superpixels

Returns

A data.frame encoding the explanations one row per explained observation. The columns are:

• model_type: The type of the model used for prediction.
• case: The case being explained (the rowname in cases).
• model_r2: The quality of the model used for the explanation
• model_intercept: The intercept of the model used for the explanation
• model_prediction: The prediction of the observation based on the model used for the explanation.
• feature: The feature used for the explanation
• feature_value: The value of the feature used
• feature_weight: The weight of the feature in the explanation
• feature_desc: A human readable description of the feature importance.
• data: Original data being explained
• prediction: The original prediction from the model

Furthermore classification explanations will also contain:

• label: The label being explained
• label_prob: The probability of label as predicted by model

Examples

# Explaining a model and an explainer for it
library(MASS)
iris_test <- iris[1, 1:4]
iris_train <- iris[-1, 1:4]
iris_lab <- iris[[5]][-1]
model <- lda(iris_train, iris_lab)
explanation <- lime(iris_train, model)

# This can now be used together with the explain method
explain(iris_test, explanation, n_labels = 1, n_features = 2)