Ensemble fixed for new SVM's

orionw · Jan 8, 2019 · 876ad80 · 876ad80
1 parent 7c4fa4c
commit 876ad80
Show file tree

Hide file tree

Showing 8 changed files with 58 additions and 17 deletions.
diff --git a/NAMESPACE b/NAMESPACE
@@ -18,6 +18,7 @@ export(GetModelComparisons)
 export(GetModelWeights)
 export(GetPredType)
 export(GetPredictionsForStacking)
+export(GetSVMScale)
 export(GetTrainingInfo)
 export(GetWeightsFromTestingSet)
 export(MajorityVote)

diff --git a/R/Ensembling.R b/R/Ensembling.R
@@ -13,7 +13,7 @@
 #' @keywords internal
 #' @export
 StripPredictions <- function(pred) {
-  if (class(pred) == "list") {
+  if (class(pred)[[1]] == "list") {
     realPredictions <- list(length = length(pred))
     i = 0
     for (model.pred in pred) {
@@ -84,7 +84,7 @@ predict.Ensemble <- function(object, newdata, voting.type="default", ...) {
   i = 0
   for (model in object$models) {
     i = i + 1
-    preds[[i]] <-  StripPredictions(predict(model, newdata = newdata, type="prob"))
+    preds[[i]] <-  StripPredictions(GetPredType(model, newdata))
   }
 
     # input overrides set voting type
@@ -288,16 +288,25 @@ GetFactorEqual <- function(pred) {
 GetWeightsFromTestingSet <- function(ensemble, df.train, test.set, train.type) {
   i = 0
   preds = list(length = length(ensemble$models))
-  for (model in ensemble$models) {
-    i = i + 1
-    preds[[i]] <- predict(model, df.train, type="prob")
-  }
+  fakeModelComp <- list()
+  class(fakeModelComp) <- "ModelComparison"
+  fakeModelComp$model.list <- ensemble$models
+  fakeModelComp$.multi.class <- FALSE
+  preds <- predict.ModelComparison(fakeModelComp, df.train, type="prob")
+  # names(preds) <- names(ensemble$models)
+
   weights <- list(length = length(ensemble$models))
   test.set <- GetFactorEqual(test.set)
   i = 0
   for (ind.pred in preds) {
     i = i + 1
-    conf.matrix = caret::confusionMatrix(test.set, as.factor(round(ind.pred[, 1])))
+    # those tricky SVM's
+    if (names(preds)[[i]] == "svm.formula") {
+      pred <- GetSVMScale(ind.pred)
+      conf.matrix = caret::confusionMatrix(test.set, as.factor(pred))
+    } else {
+      conf.matrix = caret::confusionMatrix(test.set, as.factor(round(ind.pred[, 1])))
+    }
     weights[[i]] = conf.matrix$byClass[train.type][[1]]
   }
   return(weights)

diff --git a/R/ModelGeneration.R b/R/ModelGeneration.R
@@ -162,6 +162,8 @@ predict.ModelComparison <- function(object, newdata, ...) {
       i = i + 1
       pred.basic[[i]] = GetPredType(model, newdata)
     }
+    .getClass <- function(model) { return(class(model)[[1]])    }
+    names(pred.basic) <- sapply(object$model.list, .getClass)
     return(pred.basic)
   }
 }

diff --git a/R/Visualization.R b/R/Visualization.R
@@ -54,8 +54,6 @@ plot.ModelComparison <- function(object, labels, training.data = "none", predict
   if (predictions == "empty") {
     # Predictions not given - create them here from training data
     pred.basic <- predict(object, newdata=training.data, type="prob")
-    .getClass <- function(model) { return(class(model)[[1]])    }
-    names(pred.basic) <- sapply(object$model.list, .getClass)
   } else {
     # use the given predictions
     pred.basic = predictions

diff --git a/man/GetSVMScale.Rd b/man/GetSVMScale.Rd
diff --git a/man/plot.ModelComparison.Rd b/man/plot.ModelComparison.Rd
diff --git a/tests/testthat/test_ensemble_integration.R b/tests/testthat/test_ensemble_integration.R
@@ -3,10 +3,6 @@ context("Ensemble - Model Comparison Integration")
 # load the libraries
 library(BestModel)
 
-# test_that("Ensemble plot", {
-#   # TODO: decide what to do with this
-# })
-
 test_that("Ensemble used in a ModelComparison", {
   # prepare the dataset
   iris <- PrepareIris()
@@ -21,5 +17,5 @@ test_that("Ensemble used in a ModelComparison", {
   comp <- ModelComparison(mlist, F)
   expect_equal(class(comp), "ModelComparison")
   # make sure plot works
-  print(plot(comp, iris[, 5], iris[,1:4]))
+  plot(comp, iris[, 5], iris[,1:4])
 })
diff --git a/tests/testthat/test_package_integration.R b/tests/testthat/test_package_integration.R
@@ -44,8 +44,11 @@ test_that("Other packages used for ModelComparison", {
   # multiple metrics, two word metrics, uncapitalized
   plot(comp, titanic[, 1], titanic[, -1], plot.type=list("precision", "accuracy",
                                                "recall", "detection rate"))
+  # verify we can  make an Ensemble
+  ensem <- Ensemble(comp$model.list, "majorityWeight", titanic[, -1], titanic[, 1])
+  expect_equal(length(ensem$weight.list), length(models))
 
-
-  # ensem1 <- Ensemble(comp$model.list, "majorityWeight", iris[,1:4], iris[,5])
+  pred.ensem <- predict(ensem, titanic[, -1])
+  expect_equal(length(titanic[, 1]), length(pred.ensem))
 
 })