fix percentile computation for regression objectives (#5848)

* Update regression_objective.hpp * Update regression_objective.hpp Maybe still need a (1.0 - alpha) * fix position in percentile calculation * fix regression metric threshold for l1 --------- Co-authored-by: shiyu1994 <shiyu_k1994@qq.com>
microsoft · Aug 15, 2023 · 37c3d3f · 37c3d3f
1 parent 5c9e61d
commit 37c3d3f
Show file tree

Hide file tree

Showing 2 changed files with 4 additions and 4 deletions.
diff --git a/src/objective/regression_objective.hpp b/src/objective/regression_objective.hpp
@@ -24,14 +24,14 @@ namespace LightGBM {
     for (data_size_t i = 0; i < cnt_data; ++i) {                          \
       ref_data[i] = data_reader(i);                                       \
     }                                                                     \
-    const double float_pos = static_cast<double>(1.0 - alpha) * cnt_data; \
-    const data_size_t pos = static_cast<data_size_t>(float_pos);          \
+    const double float_pos = static_cast<double>(cnt_data - 1) * (1.0 - alpha);   \
+    const data_size_t pos = static_cast<data_size_t>(float_pos) + 1;      \
     if (pos < 1) {                                                        \
       return ref_data[ArrayArgs<T>::ArgMax(ref_data)];                    \
     } else if (pos >= cnt_data) {                                         \
       return ref_data[ArrayArgs<T>::ArgMin(ref_data)];                    \
     } else {                                                              \
-      const double bias = float_pos - pos;                                \
+      const double bias = float_pos - (pos - 1);                          \
       if (pos > cnt_data / 2) {                                           \
         ArrayArgs<T>::ArgMaxAtK(&ref_data, 0, cnt_data, pos - 1);         \
         T v1 = ref_data[pos - 1];                                         \

diff --git a/tests/python_package_test/test_engine.py b/tests/python_package_test/test_engine.py
@@ -142,7 +142,7 @@ def test_regression(objective):
     elif objective == 'quantile':
         assert ret < 1311
     else:
-        assert ret < 338
+        assert ret < 343
     assert evals_result['valid_0']['l2'][-1] == pytest.approx(ret)