fix calculation of weighted gamma loss (fixes #4174)

R-package/tests/testthat/test_weighted_loss.R

@@ -0,0 +1,67 @@
+context("Case weights are respected")
+
+test_that("Gamma regression reacts on 'weight'", {
+  n <- 100L
+  set.seed(87L)
+  X <- matrix(runif(2L * n), ncol = 2L)
+  y <- X[, 1L] + X[, 2L] + runif(n)
+  X_pred <- X[1L:5L, ]
+
+  params <- list(objective = "gamma")
+
+  # Unweighted
+  dtrain <- lgb.Dataset(X, label = y)
+  bst <- lgb.train(
+    params = params
+    , data = dtrain
+    , nrounds = 4L
+    , verbose = 0L
+  )
+  pred_unweighted <- predict(bst, X_pred)
+
+  # Constant weight 1
+  dtrain <- lgb.Dataset(
+    X
+    , label = y
+    , weight = rep(1.0, n)
+  )
+  bst <- lgb.train(
+    params = params
+    , data = dtrain
+    , nrounds = 4L
+    , verbose = 0L
+  )
+  pred_weighted_1 <- predict(bst, X_pred)
+
+  # Constant weight 2
+  dtrain <- lgb.Dataset(
+    X
+    , label = y
+    , weight = rep(2.0, n)
+  )
+  bst <- lgb.train(
+    params = params
+    , data = dtrain
+    , nrounds = 4L
+    , verbose = 0L
+  )
+  pred_weighted_2 <- predict(bst, X_pred)
+
+  # Non-constant weights
+  dtrain <- lgb.Dataset(
+    X
+    , label = y
+    , weight = seq(0.0, 1.0, length.out = n)
+  )
+  bst <- lgb.train(
+    params = params
+    , data = dtrain
+    , nrounds = 4L
+    , verbose = 0L
+  )
+  pred_weighted <- predict(bst, X_pred)
+
+  expect_equal(pred_unweighted, pred_weighted_1)
+  expect_equal(pred_weighted_1, pred_weighted_2)
+  expect_false(all(pred_unweighted == pred_weighted))
+})

src/objective/regression_objective.hpp

@@ -695,7 +695,7 @@ class RegressionGammaLoss : public RegressionPoissonLoss {
     } else {
       #pragma omp parallel for schedule(static)
       for (data_size_t i = 0; i < num_data_; ++i) {
-        gradients[i] = static_cast<score_t>(1.0 - label_[i] * std::exp(-score[i]) * weights_[i]);
+        gradients[i] = static_cast<score_t>((1.0 - label_[i] * std::exp(-score[i])) * weights_[i]);
         hessians[i] = static_cast<score_t>(label_[i] * std::exp(-score[i]) * weights_[i]);
       }
     }