[Breaking] Change default evaluation metric for classification to log…

…loss / mlogloss (#6183)

* Change DefaultEvalMetric of classification from error to logloss

* Change default binary metric in plugin/example/custom_obj.cc

* Set old error metric in python tests

* Set old error metric in R tests

* Fix missed eval metrics and typos in R tests

* Fix setting eval_metric twice in R tests

* Add warning for empty eval_metric for classification

* Fix Dask tests

Co-authored-by: Hyunsu Cho <chohyu01@cs.washington.edu>

R-package/tests/testthat/test_basic.R

@@ -17,7 +17,8 @@ test_that("train and predict binary classification", {
   nrounds <- 2
   expect_output(
     bst <- xgboost(data = train$data, label = train$label, max_depth = 2,
-                  eta = 1, nthread = 2, nrounds = nrounds, objective = "binary:logistic")
+                  eta = 1, nthread = 2, nrounds = nrounds, objective = "binary:logistic",
+                  eval_metric = "error")
   , "train-error")
   expect_equal(class(bst), "xgb.Booster")
   expect_equal(bst$niter, nrounds)
@@ -122,7 +123,7 @@ test_that("train and predict softprob", {
   expect_output(
     bst <- xgboost(data = as.matrix(iris[, -5]), label = lb,
                    max_depth = 3, eta = 0.5, nthread = 2, nrounds = 5,
-                   objective = "multi:softprob", num_class = 3)
+                   objective = "multi:softprob", num_class = 3, eval_metric = "merror")
   , "train-merror")
   expect_false(is.null(bst$evaluation_log))
   expect_lt(bst$evaluation_log[, min(train_merror)], 0.025)
@@ -150,7 +151,7 @@ test_that("train and predict softmax", {
   expect_output(
     bst <- xgboost(data = as.matrix(iris[, -5]), label = lb,
                    max_depth = 3, eta = 0.5, nthread = 2, nrounds = 5,
-                   objective = "multi:softmax", num_class = 3)
+                   objective = "multi:softmax", num_class = 3, eval_metric = "merror")
   , "train-merror")
   expect_false(is.null(bst$evaluation_log))
   expect_lt(bst$evaluation_log[, min(train_merror)], 0.025)
@@ -167,7 +168,7 @@ test_that("train and predict RF", {
   lb <- train$label
   # single iteration
   bst <- xgboost(data = train$data, label = lb, max_depth = 5,
-                 nthread = 2, nrounds = 1, objective = "binary:logistic",
+                 nthread = 2, nrounds = 1, objective = "binary:logistic", eval_metric = "error",
                  num_parallel_tree = 20, subsample = 0.6, colsample_bytree = 0.1)
   expect_equal(bst$niter, 1)
   expect_equal(xgb.ntree(bst), 20)
@@ -193,7 +194,8 @@ test_that("train and predict RF with softprob", {
   set.seed(11)
   bst <- xgboost(data = as.matrix(iris[, -5]), label = lb,
                  max_depth = 3, eta = 0.9, nthread = 2, nrounds = nrounds,
-                 objective = "multi:softprob", num_class = 3, verbose = 0,
+                 objective = "multi:softprob", eval_metric = "merror",
+                 num_class = 3, verbose = 0,
                  num_parallel_tree = 4, subsample = 0.5, colsample_bytree = 0.5)
   expect_equal(bst$niter, 15)
   expect_equal(xgb.ntree(bst), 15 * 3 * 4)
@@ -274,7 +276,7 @@ test_that("xgb.cv works", {
   expect_output(
     cv <- xgb.cv(data = train$data, label = train$label, max_depth = 2, nfold = 5,
                  eta = 1., nthread = 2, nrounds = 2, objective = "binary:logistic",
-                 verbose = TRUE)
+                 eval_metric = "error", verbose = TRUE)
   , "train-error:")
   expect_is(cv, 'xgb.cv.synchronous')
   expect_false(is.null(cv$evaluation_log))
@@ -299,7 +301,7 @@ test_that("xgb.cv works with stratified folds", {
                 eta = 1., nthread = 2, nrounds = 2, objective = "binary:logistic",
                 verbose = TRUE, stratified = TRUE)
   # Stratified folds should result in a different evaluation logs
-  expect_true(all(cv$evaluation_log[, test_error_mean] != cv2$evaluation_log[, test_error_mean]))
+  expect_true(all(cv$evaluation_log[, test_logloss_mean] != cv2$evaluation_log[, test_logloss_mean]))
 })
 
 test_that("train and predict with non-strict classes", {

R-package/tests/testthat/test_callbacks.R

@@ -26,7 +26,8 @@ watchlist <- list(train = dtrain, test = dtest)
 
 err <- function(label, pr) sum((pr > 0.5) != label) / length(label)
 
-param <- list(objective = "binary:logistic", max_depth = 2, nthread = 2)
+param <- list(objective = "binary:logistic", eval_metric = "error",
+              max_depth = 2, nthread = 2)
 
 
 test_that("cb.print.evaluation works as expected", {
@@ -105,7 +106,8 @@ test_that("cb.evaluation.log works as expected", {
 })
 
 
-param <- list(objective = "binary:logistic", max_depth = 4, nthread = 2)
+param <- list(objective = "binary:logistic", eval_metric = "error",
+              max_depth = 4, nthread = 2)
 
 test_that("can store evaluation_log without printing", {
   expect_silent(
@@ -236,7 +238,7 @@ test_that("early stopping xgb.train works", {
 test_that("early stopping using a specific metric works", {
   set.seed(11)
   expect_output(
-    bst <- xgb.train(param, dtrain, nrounds = 20, watchlist, eta = 0.6,
+    bst <- xgb.train(param[-2], dtrain, nrounds = 20, watchlist, eta = 0.6,
                      eval_metric = "logloss", eval_metric = "auc",
                      callbacks = list(cb.early.stop(stopping_rounds = 3, maximize = FALSE,
                                                     metric_name = 'test_logloss')))

R-package/tests/testthat/test_glm.R

@@ -8,7 +8,7 @@ test_that("gblinear works", {
   dtrain <- xgb.DMatrix(agaricus.train$data, label = agaricus.train$label)
   dtest <- xgb.DMatrix(agaricus.test$data, label = agaricus.test$label)
 
-  param <- list(objective = "binary:logistic", booster = "gblinear",
+  param <- list(objective = "binary:logistic", eval_metric = "error", booster = "gblinear",
                 nthread = 2, eta = 0.8, alpha = 0.0001, lambda = 0.0001)
   watchlist <- list(eval = dtest, train = dtrain)
 

demo/guide-python/custom_softmax.py

@@ -142,7 +142,8 @@ def main(args):
 
     native_results = {}
     # Use the same objective function defined in XGBoost.
-    booster_native = xgb.train({'num_class': kClasses},
+    booster_native = xgb.train({'num_class': kClasses,
+                                'eval_metric': 'merror'},
                                m,
                                num_boost_round=kRounds,
                                evals_result=native_results,

doc/parameter.rst

@@ -376,7 +376,7 @@ Specify the learning task and the corresponding learning objective. The objectiv
 
 * ``eval_metric`` [default according to objective]
 
-  - Evaluation metrics for validation data, a default metric will be assigned according to objective (rmse for regression, and error for classification, mean average precision for ranking)
+  - Evaluation metrics for validation data, a default metric will be assigned according to objective (rmse for regression, and logloss for classification, mean average precision for ranking)
   - User can add multiple evaluation metrics. Python users: remember to pass the metrics in as list of parameters pairs instead of map, so that latter ``eval_metric`` won't override previous one
   - The choices are listed below:
 

jvm-packages/xgboost4j-spark/src/main/scala/ml/dmlc/xgboost4j/scala/spark/XGBoostClassifier.scala

@@ -154,10 +154,10 @@ class XGBoostClassifier (
     require(isDefined(objective), "Users must set \'objective\' via xgboostParams.")
     if ($(objective).startsWith("multi")) {
       // multi
-      "merror"
+      "mlogloss"
     } else {
       // binary
-      "error"
+      "logloss"
     }
   }
 

plugin/example/custom_obj.cc

@@ -56,7 +56,7 @@ class MyLogistic : public ObjFunction {
     }
   }
   const char* DefaultEvalMetric() const override {
-    return "error";
+    return "logloss";
   }
   void PredTransform(HostDeviceVector<bst_float> *io_preds) override {
     // transform margin value to probability.

plugin/updater_oneapi/regression_loss_oneapi.h

@@ -103,7 +103,7 @@ struct LogisticRegressionOneAPI {
 
 // logistic loss for binary classification task
 struct LogisticClassificationOneAPI : public LogisticRegressionOneAPI {
-  static const char* DefaultEvalMetric() { return "error"; }
+  static const char* DefaultEvalMetric() { return "logloss"; }
   static const char* Name() { return "binary:logistic_oneapi"; }
 };
 

src/learner.cc

@@ -1031,6 +1031,18 @@ class LearnerImpl : public LearnerIO {
     std::ostringstream os;
     os << '[' << iter << ']' << std::setiosflags(std::ios::fixed);
     if (metrics_.size() == 0 && tparam_.disable_default_eval_metric <= 0) {
+      auto warn_default_eval_metric = [](const std::string& objective, const std::string& before,
+                                         const std::string& after) {
+        LOG(WARNING) << "Starting in XGBoost 1.3.0, the default evaluation metric used with the "
+                     << "objective '" << objective << "' was changed from '" << before
+                     << "' to '" << after << "'. Explicitly set eval_metric if you'd like to "
+                     << "restore the old behavior.";
+      };
+      if (tparam_.objective == "binary:logistic") {
+        warn_default_eval_metric(tparam_.objective, "error", "logloss");
+      } else if ((tparam_.objective == "multi:softmax" || tparam_.objective == "multi:softprob")) {
+        warn_default_eval_metric(tparam_.objective, "merror", "mlogloss");
+      }
       metrics_.emplace_back(Metric::Create(obj_->DefaultEvalMetric(), &generic_parameters_));
       metrics_.back()->Configure({cfg_.begin(), cfg_.end()});
     }

src/objective/multiclass_obj.cu

@@ -125,7 +125,7 @@ class SoftmaxMultiClassObj : public ObjFunction {
     this->Transform(io_preds, true);
   }
   const char* DefaultEvalMetric() const override {
-    return "merror";
+    return "mlogloss";
   }
 
   inline void Transform(HostDeviceVector<bst_float> *io_preds, bool prob) {

src/objective/regression_loss.h

@@ -131,7 +131,7 @@ struct PseudoHuberError {
 
 // logistic loss for binary classification task
 struct LogisticClassification : public LogisticRegression {
-  static const char* DefaultEvalMetric() { return "error"; }
+  static const char* DefaultEvalMetric() { return "logloss"; }
   static const char* Name() { return "binary:logistic"; }
 };
 

tests/cpp/plugin/test_example_objective.cc

@@ -8,7 +8,7 @@ namespace xgboost {
 TEST(Plugin, ExampleObjective) {
   xgboost::GenericParameter tparam = CreateEmptyGenericParam(GPUIDX);
   auto * obj = xgboost::ObjFunction::Create("mylogistic", &tparam);
-  ASSERT_EQ(obj->DefaultEvalMetric(), std::string{"error"});
+  ASSERT_EQ(obj->DefaultEvalMetric(), std::string{"logloss"});
   delete obj;
 }
 

tests/python/test_basic.py

@@ -81,7 +81,7 @@ def test_record_results(self):
         dtrain = xgb.DMatrix(dpath + 'agaricus.txt.train')
         dtest = xgb.DMatrix(dpath + 'agaricus.txt.test')
         param = {'max_depth': 2, 'eta': 1, 'verbosity': 0,
-                 'objective': 'binary:logistic'}
+                 'objective': 'binary:logistic', 'eval_metric': 'error'}
         # specify validations set to watch performance
         watchlist = [(dtest, 'eval'), (dtrain, 'train')]
         num_round = 2

tests/python/test_basic_models.py

@@ -117,7 +117,8 @@ def run_eta_decay(self, tree_method):
         # learning_rates as a list
         # init eta with 0 to check whether learning_rates work
         param = {'max_depth': 2, 'eta': 0, 'verbosity': 0,
-                 'objective': 'binary:logistic', 'tree_method': tree_method}
+                 'objective': 'binary:logistic', 'eval_metric': 'error',
+                 'tree_method': tree_method}
         evals_result = {}
         bst = xgb.train(param, dtrain, num_round, watchlist,
                         callbacks=[xgb.callback.reset_learning_rate([
@@ -131,7 +132,8 @@ def run_eta_decay(self, tree_method):
 
         # init learning_rate with 0 to check whether learning_rates work
         param = {'max_depth': 2, 'learning_rate': 0, 'verbosity': 0,
-                 'objective': 'binary:logistic', 'tree_method': tree_method}
+                 'objective': 'binary:logistic', 'eval_metric': 'error',
+                 'tree_method': tree_method}
         evals_result = {}
         bst = xgb.train(param, dtrain, num_round, watchlist,
                         callbacks=[xgb.callback.reset_learning_rate(
@@ -145,7 +147,7 @@ def run_eta_decay(self, tree_method):
         # check if learning_rates override default value of eta/learning_rate
         param = {
             'max_depth': 2, 'verbosity': 0, 'objective': 'binary:logistic',
-            'tree_method': tree_method
+            'eval_metric': 'error', 'tree_method': tree_method
         }
         evals_result = {}
         bst = xgb.train(param, dtrain, num_round, watchlist,

tests/python/test_dmatrix.py

@@ -115,7 +115,9 @@ def test_slice(self):
 
         eval_res_0 = {}
         booster = xgb.train(
-            {'num_class': 3, 'objective': 'multi:softprob'}, d,
+            {'num_class': 3, 'objective': 'multi:softprob',
+             'eval_metric': 'merror'},
+            d,
             num_boost_round=2, evals=[(d, 'd')], evals_result=eval_res_0)
 
         predt = booster.predict(d)
@@ -130,9 +132,11 @@ def test_slice(self):
         assert sliced_margin.shape[0] == len(ridxs) * 3
 
         eval_res_1 = {}
-        xgb.train({'num_class': 3, 'objective': 'multi:softprob'}, sliced,
-                  num_boost_round=2, evals=[(sliced, 'd')],
-                  evals_result=eval_res_1)
+        xgb.train(
+            {'num_class': 3, 'objective': 'multi:softprob',
+             'eval_metric': 'merror'},
+            sliced,
+            num_boost_round=2, evals=[(sliced, 'd')], evals_result=eval_res_1)
 
         eval_res_0 = eval_res_0['d']['merror']
         eval_res_1 = eval_res_1['d']['merror']

tests/python/test_early_stopping.py

@@ -58,7 +58,7 @@ def test_cv_early_stopping(self):
         y = digits['target']
         dm = xgb.DMatrix(X, label=y)
         params = {'max_depth': 2, 'eta': 1, 'verbosity': 0,
-                  'objective': 'binary:logistic'}
+                  'objective': 'binary:logistic', 'eval_metric': 'error'}
 
         cv = xgb.cv(params, dm, num_boost_round=10, nfold=10,
                     early_stopping_rounds=10)

tests/python/test_with_dask.py

@@ -274,7 +274,7 @@ def test_dask_classifier():
             X, y = generate_array()
             y = (y * 10).astype(np.int32)
             classifier = xgb.dask.DaskXGBClassifier(
-                verbosity=1, n_estimators=2)
+                verbosity=1, n_estimators=2, eval_metric='merror')
             classifier.client = client
             classifier.fit(X, y, eval_set=[(X, y)])
             prediction = classifier.predict(X)
@@ -386,6 +386,7 @@ def _check_outputs(out, predictions):
     y = dd.from_array(np.random.randint(low=0, high=n_classes, size=kRows))
     dtrain = xgb.dask.DaskDMatrix(client, X, y)
     parameters['objective'] = 'multi:softprob'
+    parameters['eval_metric'] = 'merror'
     parameters['num_class'] = n_classes
 
     out = xgb.dask.train(client, parameters,
@@ -482,7 +483,7 @@ async def run_dask_classifier_asyncio(scheduler_address):
         X, y = generate_array()
         y = (y * 10).astype(np.int32)
         classifier = await xgb.dask.DaskXGBClassifier(
-            verbosity=1, n_estimators=2)
+            verbosity=1, n_estimators=2, eval_metric='merror')
         classifier.client = client
         await classifier.fit(X, y, eval_set=[(X, y)])
         prediction = await classifier.predict(X)

tests/python/test_with_pandas.py

@@ -174,7 +174,7 @@ def test_pandas_weight(self):
     def test_cv_as_pandas(self):
         dm = xgb.DMatrix(dpath + 'agaricus.txt.train')
         params = {'max_depth': 2, 'eta': 1, 'verbosity': 0,
-                  'objective': 'binary:logistic'}
+                  'objective': 'binary:logistic', 'eval_metric': 'error'}
 
         cv = xgb.cv(params, dm, num_boost_round=10, nfold=10)
         assert isinstance(cv, pd.DataFrame)