glmnet: importance() method

.github/workflows/R-CMD-check-extra.yaml

@@ -1,9 +1,4 @@
-on:
-  push:
-    branches:
-      - master
-  schedule:
-    - cron: '15 8 * * *'
+on: [push, pull_request]
 
 name: R-CMD-check-extra
 

.github/workflows/R-CMD-check.yaml

@@ -1,10 +1,4 @@
-on:
-  push:
-    branches:
-      - master
-  pull_request:
-    branches:
-      - master
+on: [push, pull_request]
 
 name: R-CMD-check
 

.gitignore

@@ -24,4 +24,5 @@
 .DS_Store
 .Rproj.user
 
-docs/
+docs/
+.vscode

DESCRIPTION

@@ -57,4 +57,4 @@ RdMacros: mlr3misc
 Encoding: UTF-8
 NeedsCompilation: no
 Roxygen: list(markdown = TRUE)
-RoxygenNote: 7.0.2
+RoxygenNote: 7.1.0

R/LearnerClassifGlmnet.R

@@ -60,14 +60,43 @@ LearnerClassifGlmnet = R6Class("LearnerClassifGlmnet", inherit = LearnerClassif,
         param_set = ps,
         predict_types = c("response", "prob"),
         feature_types = c("logical", "integer", "numeric"),
-        properties = c("weights", "twoclass", "multiclass"),
+        properties = c("weights", "twoclass", "multiclass", "importance"),
         packages = "glmnet",
         man = "mlr3learners::mlr_learners_classif.glmnet"
       )
+    },
+
+    #' @description
+    #' The importance scores are extracted from the beta coefficients of the
+    #' fitted model.
+    #' @param task ([Task])\cr
+    #'   Task to calculate the importance on. If a model was already fitted,
+    #' no refitting is done.
+    #' @return Named `numeric()`.
+    importance = function(task) {
+
+      if (is.null(self$model)) {
+        self$train(task)
+      }
+
+      if ("multiclass" %in% task$properties) {
+        mlr3misc::stopf("Learner 'glmnet' only support feature importance for
+          'twoclass' tasks.", wrap = TRUE)
+      }
+      model = self$model$glmnet.fit
+
+      res = mlr3misc::map_dbl(seq_len(nrow(model$beta)), function(.i) {
+        ind = which(model$beta[.i, ] != 0)[1]
+        model$lambda[ind]
+      })
+
+      names(res) = model$beta@Dimnames[[1]]
+      sort(res, decreasing = TRUE)
     }
   ),
 
   private = list(
+
     .train = function(task) {
 
       pars = self$param_set$get_values(tags = "train")
@@ -76,7 +105,7 @@ LearnerClassifGlmnet = R6Class("LearnerClassifGlmnet", inherit = LearnerClassif,
       if ("weights" %in% task$properties) {
         pars$weights = task$weights$weight
       }
-      pars$family = ifelse(length(task$class_names) == 2L, "binomial", "multinomial")
+      pars$family = ifelse(length(task$class_names) == 2L, "binomial", "multinomial") # nolint
 
       saved_ctrl = glmnet::glmnet.control()
       on.exit(invoke(glmnet::glmnet.control, .args = saved_ctrl))
@@ -96,10 +125,12 @@ LearnerClassifGlmnet = R6Class("LearnerClassifGlmnet", inherit = LearnerClassif,
       newdata = as.matrix(task$data(cols = task$feature_names))
 
       if (self$predict_type == "response") {
-        response = invoke(predict, self$model, newx = newdata, type = "class", .args = pars)
+        response = invoke(predict, self$model, newx = newdata, type = "class",
+          .args = pars)
         PredictionClassif$new(task = task, response = drop(response))
       } else {
-        prob = invoke(predict, self$model, newx = newdata, type = "response", .args = pars)
+        prob = invoke(predict, self$model, newx = newdata, type = "response",
+          .args = pars)
 
         if (length(task$class_names) == 2L) {
           prob = cbind(prob, 1 - prob)

R/LearnerRegrGlmnet.R

@@ -67,10 +67,31 @@ LearnerRegrGlmnet = R6Class("LearnerRegrGlmnet", inherit = LearnerRegr,
         id = "regr.glmnet",
         param_set = ps,
         feature_types = c("logical", "integer", "numeric"),
-        properties = "weights",
+        properties = c("weights", "importance"),
         packages = "glmnet",
         man = "mlr3learners::mlr_learners_regr.glmnet"
       )
+    },
+
+    #' @description
+    #' The importance scores are extracted from the beta coefficients of the
+    #' fitted model.
+    #' @return Named `numeric()`.
+    importance = function() {
+
+      if (is.null(self$model)) {
+        stopf("No model stored. Please train the model first.")
+      }
+
+      model = self$model$glmnet.fit
+
+      res = mlr3misc::map_dbl(seq_len(nrow(model$beta)), function(.i) {
+        ind = which(model$beta[.i, ] != 0)[1]
+        model$lambda[ind]
+      })
+
+      names(res) = model$beta@Dimnames[[1]]
+      sort(res, decreasing = TRUE)
     }
   ),
 
@@ -100,7 +121,8 @@ LearnerRegrGlmnet = R6Class("LearnerRegrGlmnet", inherit = LearnerRegr,
       pars = self$param_set$get_values(tags = "predict")
       newdata = as.matrix(task$data(cols = task$feature_names))
 
-      response = invoke(predict, self$model, newx = newdata, type = "response", .args = pars)
+      response = invoke(predict, self$model, newx = newdata, type = "response",
+       .args = pars)
       PredictionRegr$new(task = task, response = drop(response))
     }
   )