average_loss: Average Loss

In hstats: Interaction Statistics

Average Loss

Description

Calculates the average loss of a model on a given dataset, optionally grouped by a variable. Use plot() to visualize the results.

Usage

average_loss(object, ...)

## Default S3 method:
average_loss(
  object,
  X,
  y,
  pred_fun = stats::predict,
  loss = "squared_error",
  agg_cols = FALSE,
  BY = NULL,
  by_size = 4L,
  w = NULL,
  ...
)

## S3 method for class 'ranger'
average_loss(
  object,
  X,
  y,
  pred_fun = function(m, X, ...) stats::predict(m, X, ...)$predictions,
  loss = "squared_error",
  agg_cols = FALSE,
  BY = NULL,
  by_size = 4L,
  w = NULL,
  ...
)

## S3 method for class 'explainer'
average_loss(
  object,
  X = object[["data"]],
  y = object[["y"]],
  pred_fun = object[["predict_function"]],
  loss = "squared_error",
  agg_cols = FALSE,
  BY = NULL,
  by_size = 4L,
  w = object[["weights"]],
  ...
)

Arguments

object	Fitted model object.
...	Additional arguments passed to pred_fun(object, X, ...), for instance type = "response" in a glm() model, or reshape = TRUE in a multiclass XGBoost model.
X	A data.frame or matrix serving as background dataset.
y	Vector/matrix of the response, or the corresponding column names in X.
pred_fun	Prediction function of the form ⁠function(object, X, ...)⁠, providing K \ge 1 predictions per row. Its first argument represents the model object, its second argument a data structure like X. Additional arguments (such as type = "response" in a GLM, or reshape = TRUE in a multiclass XGBoost model) can be passed via .... The default, stats::predict(), will work in most cases.
loss	One of "squared_error", "logloss", "mlogloss", "poisson", "gamma", "absolute_error", "classification_error". Alternatively, a loss function can be provided that turns observed and predicted values into a numeric vector or matrix of unit losses of the same length as X. For "mlogloss", the response y can either be a dummy matrix or a discrete vector. The latter case is handled via a fast version of model.matrix(~ as.factor(y) + 0). For "classification_error", both predictions and responses can be non-numeric. For "squared_error", both predictions and responses can be factors with identical levels. In this case, squared error is evaulated for each one-hot-encoded column.
agg_cols	Should multivariate losses be summed up? Default is FALSE. In combination with the squared error loss, agg_cols = TRUE gives the Brier score for (probabilistic) classification.
BY	Optional grouping vector or column name. Numeric BY variables with more than by_size disjoint values will be binned into by_size quantile groups of similar size.
by_size	Numeric BY variables with more than by_size unique values will be binned into quantile groups. Only relevant if BY is not NULL.
w	Optional vector of case weights. Can also be a column name of X.

Value

An object of class "hstats_matrix" containing these elements:

• M: Matrix of statistics (one column per prediction dimension), or NULL.

• SE: Matrix with standard errors of M, or NULL. Multiply with sqrt(m_rep) to get standard deviations instead. Currently, supported only for perm_importance().

• m_rep: The number of repetitions behind standard errors SE, or NULL. Currently, supported only for perm_importance().

• statistic: Name of the function that generated the statistic.

• description: Description of the statistic.

Methods (by class)

• average_loss(default): Default method.

• average_loss(ranger): Method for "ranger" models.

• average_loss(explainer): Method for DALEX "explainer".

Losses

The default loss is the "squared_error". Other choices:

• "absolute_error": The absolute error is the loss corresponding to median regression.

• "poisson": Unit Poisson deviance, i.e., the loss function used in Poisson regression. Actual values y and predictions must be non-negative.

• "gamma": Unit gamma deviance, i.e., the loss function of Gamma regression. Actual values y and predictions must be positive.

• "logloss": The Log Loss is the loss function used in logistic regression, and the top choice in probabilistic binary classification. Responses y and predictions must be between 0 and 1. Predictions represent probabilities of having a "1".

• "mlogloss": Multi-Log-Loss is the natural loss function in probabilistic multi-class situations. If there are K classes and n observations, the predictions form a (n x K) matrix of probabilities (with row-sums 1). The observed values y are either passed as (n x K) dummy matrix, or as discrete vector with corresponding levels. The latter case is turned into a dummy matrix by a fast version of model.matrix(~ as.factor(y) + 0).

• "classification_error": Misclassification error. Both the observed values y and the predictions can be character/factor. This loss function can be used in non-probabilistic classification settings. BUT: Probabilistic classification (with "mlogloss") is clearly preferred in most situations.

• A function with signature f(actual, predicted), returning a numeric vector or matrix of the same length as the input.

Examples

# MODEL 1: Linear regression
fit <- lm(Sepal.Length ~ ., data = iris)
average_loss(fit, X = iris, y = "Sepal.Length")
average_loss(fit, X = iris, y = iris$Sepal.Length, BY = iris$Sepal.Width)
average_loss(fit, X = iris, y = "Sepal.Length", BY = "Sepal.Width")

# MODEL 2: Multi-response linear regression
fit <- lm(as.matrix(iris[, 1:2]) ~ Petal.Length + Petal.Width + Species, data = iris)
average_loss(fit, X = iris, y = iris[, 1:2])
L <- average_loss(
  fit, X = iris, y = iris[, 1:2], loss = "gamma", BY = "Species"
)
L
plot(L)

hstats documentation built on May 29, 2024, 6:43 a.m.