feat: support for linex loss (#49)

.lintr

@@ -1,10 +1,10 @@
 linters: linters_with_defaults(
-    # lintr defaults: https://github.com/jimhester/lintr#available-linters
+    # lintr defaults: https://lintr.r-lib.org/reference/default_linters.html
     # the following setup changes/removes certain linters
     assignment_linter = NULL, # do not force using <- for assignments
     object_name_linter = object_name_linter(c("snake_case", "CamelCase")), # only allow snake case and camel case object names
     cyclocomp_linter = NULL, # do not check function complexity
     commented_code_linter = NULL, # allow code in comments
-    line_length_linter = line_length_linter(120)
+    line_length_linter = line_length_linter(120L)
     )
 

DESCRIPTION

@@ -66,6 +66,7 @@ Collate:
     'regr_ape.R'
     'regr_bias.R'
     'regr_ktau.R'
+    'regr_linex.R'
     'regr_mae.R'
     'regr_mape.R'
     'regr_maxae.R'

NAMESPACE

@@ -22,6 +22,7 @@ export(gmean)
 export(gpr)
 export(jaccard)
 export(ktau)
+export(linex)
 export(logloss)
 export(mae)
 export(mape)

NEWS.md

@@ -1,5 +1,7 @@
 # mlr3measures (development version)
 
+* Added new measure `linex` (Linear-Exponential Loss).
+
 # mlr3measures 0.6.0
 
 * Added binary classification measures `gmean` and `gpr`.

R/bibentries.R

@@ -182,6 +182,15 @@ bibentries = c( # nolint start
     pages            = "1263--1284",
     year             = "2009",
     doi              = "10.1109/TKDE.2008.239"
+  ),
+  varian_1975 = bibentry("incollection",
+    title        = "A Bayesian Approach to Real Estate Assessment",
+    author       = "Varian, Hal R.",
+    booktitle    = "Studies in Bayesian Econometrics and Statistics: In Honor of Leonard J. Savage",
+    editor       = "Stephen E. Fienberg and Arnold Zellner",
+    year         = "1975",
+    publisher    = "North-Holland",
+    address      = "Amsterdam",
+    pages        = "195--208"
   )
-
 ) # nolint end

R/binary_fbeta.R

@@ -22,7 +22,7 @@
 #' @inheritParams binary_params
 #' @param beta (`numeric(1)`)\cr
 #'   Parameter to give either precision or recall more weight.
-#'   Default is 1, resulting in balanced weights.
+#'   Default is `1`, resulting in balanced weights.
 #' @template binary_example
 #' @export
 fbeta = function(truth, response, positive, beta = 1, na_value = NaN, ...) {

R/classif_mbrier.R

@@ -7,7 +7,7 @@
 #     1/n * sum_i sum_j (I_ij - p_ij)^2.
 #' }
 #' \eqn{I_{ij}}{I_ij} is 1 if observation \eqn{x_i} has true label \eqn{j}, and 0 otherwise.
-#' \eqn{p_{ij}{p_ij} is the probability that observation \eqn{x_i} belongs to class \eqn{j}.
+#' \eqn{p_{ij}}{p_ij} is the probability that observation \eqn{x_i} belongs to class \eqn{j}.
 #'
 #' Note that there also is the more common definition of the Brier score for binary
 #' classification problems in [bbrier()].

R/classif_mcc.R

@@ -35,7 +35,7 @@
 #' `r format_bib("matthews_1975")`
 #'
 #' @inheritParams classif_params
-#' @param positive (`character(1`) Name of the positive class in case of binary classification.
+#' @param positive (`character(1)`) Name of the positive class in case of binary classification.
 #'
 #' @template classif_example
 #' @export

R/regr_linex.R

@@ -0,0 +1,40 @@
+#' @title Linear-Exponential Loss (per observation)
+#'
+#' @details
+#' The Linear-Exponential Loss is defined as \deqn{
+#'   b ( \exp (t_i - r_i) - a (t_i - r_i) - 1 ),
+#' }{
+#    b(exp(t - r) − a(t - r) − 1),
+#' },
+#' where \eqn{a \neq 0, b > 0}.
+#'
+#' @templateVar mid linex
+#' @template regr_template
+#' @param a (`numeric(1)`)\cr
+#'   Shape parameter controlling asymmetry.
+#'   Negative values penalize overestimation more, positive values penalize underestimation more.
+#'   As `a` approaches 0, the loss resembles squared error loss. Default is `-1`.
+#' @param b (`numeric(1)`)\cr
+#'   Positive scaling factor for the loss. Larger values increase the loss magnitude.
+#'   Default is `1`.
+#'
+#' @references
+#' `r format_bib("varian_1975")`
+#'
+#' @inheritParams regr_params
+#' @template regr_example
+#' @export
+linex = function(truth, response, a = -1, b = 1, ...) {
+  assert_regr(truth, response = response)
+  assert_number(a)
+  assert_number(b, lower = 0)
+  if (a == 0) {
+    stop("'a' can't be 0")
+  }
+  e = truth - response
+  we = a * e
+  b * (exp(we) - we - 1)
+}
+
+#' @include measures.R
+add_measure(linex, "Linear-exponential Loss (per observation)", "regr", 0, Inf, TRUE, aggregated = FALSE)