[backport] Fix NDCG metric with non-exp gain. (dmlc#11534) (dmlc#11578)

Author: trivialfis

src/metric/rank_metric.cc

@@ -1,5 +1,5 @@
 /**
- * Copyright 2020-2023 by XGBoost contributors
+ * Copyright 2020-2025, XGBoost contributors
  */
 #include "rank_metric.h"
 
@@ -10,7 +10,6 @@
 #include <array>                             // for array
 #include <cmath>                             // for log, sqrt
 #include <functional>                        // for less, greater
-#include <limits>                            // for numeric_limits
 #include <map>                               // for operator!=, _Rb_tree_const_iterator
 #include <memory>                            // for allocator, unique_ptr, shared_ptr, __shared_...
 #include <numeric>                           // for accumulate
@@ -22,7 +21,6 @@
 #include "../collective/aggregator.h"        // for ApplyWithLabels
 #include "../common/algorithm.h"             // for ArgSort, Sort
 #include "../common/linalg_op.h"             // for cbegin, cend
-#include "../common/math.h"                  // for CmpFirst
 #include "../common/optional_weight.h"       // for OptionalWeights, MakeOptionalWeights
 #include "dmlc/common.h"                     // for OMPException
 #include "metric_common.h"                   // for MetricNoCache, GPUMetric, PackedReduceResult
@@ -250,10 +248,6 @@ class EvalRankWithCache : public Metric {
     }
     param_.UpdateAllowUnknown(Args{});
   }
-  void Configure(Args const&) override {
-    // do not configure, otherwise the ndcg param will be forced into the same as the one in
-    // objective.
-  }
   void LoadConfig(Json const& in) override {
     if (IsA<Null>(in)) {
       return;
@@ -365,6 +359,18 @@ class EvalNDCG : public EvalRankWithCache<ltr::NDCGCache> {
  public:
   using EvalRankWithCache::EvalRankWithCache;
 
+  void Configure(Args const& args) override {
+    // do not configure, otherwise the ndcg param like top-k will be forced into the same
+    // as the one in objective. The metric has its own syntax for parameter.
+    for (auto const& [key, value] : args) {
+      // Make a special case for the exp gain parameter, which is not exposed in the
+      // metric configuration syntax.
+      if (key == "ndcg_exp_gain") {
+        this->param_.UpdateAllowUnknown(Args{{key, value}});
+      }
+    }
+  }
+
   double Eval(HostDeviceVector<float> const& preds, MetaInfo const& info,
               std::shared_ptr<ltr::NDCGCache> p_cache) override {
     if (ctx_->IsCUDA()) {

tests/cpp/common/test_parameter.cc

@@ -1,11 +1,12 @@
-/*!
- * Copyright (c) by Contributors 2019
+/**
+ * Copyright 2019-2025, XGBoost contributors
  */
 #include <gtest/gtest.h>
-
 #include <xgboost/base.h>
 #include <xgboost/parameter.h>
 
+#include "xgboost/json.h"  // for ToJson, FromJson
+
 enum class Foo : int {
   kBar = 0, kFrog = 1, kCat = 2, kDog = 3
 };
@@ -103,3 +104,16 @@ TEST(XGBoostParameter, Update) {
   a.UpdateAllowUnknown(xgboost::Args{{"f", "2.71828"}});
   ASSERT_NE(a.f, b.f);
 }
+namespace xgboost {
+TEST(XGBoostParameter, Json) {
+  UpdatableParam a, b;
+  a.UpdateAllowUnknown(Args{{"f", "1024"}, {"d", "2048"}});
+  auto ja = Json{ToJson(a)};
+
+  UpdatableParam c;
+  FromJson(ja, &c);
+  ASSERT_FLOAT_EQ(a.f, 1024);
+  ASSERT_FLOAT_EQ(c.f, 1024);
+  ASSERT_FLOAT_EQ(b.f, 0);  // Make sure dmlc global variable is not used here.
+}
+}  // namespace xgboost

tests/python/test_ranking.py

@@ -69,6 +69,28 @@ def ndcg_gain(y: np.ndarray) -> np.ndarray:
     )
 
 
+def test_ndcg_non_exp() -> None:
+    # NDCG exp gain must have label smaller than 32
+    X, y, q, w = tm.make_ltr(n_samples=1024, n_features=4, n_query_groups=3, max_rel=44)
+
+    def fit(ltr: xgboost.XGBRanker):
+        ltr.fit(
+            X,
+            y,
+            qid=q,
+            sample_weight=w,
+            eval_set=[(X, y)],
+            eval_qid=(q,),
+            sample_weight_eval_set=(w,),
+        )
+
+    ltr = xgboost.XGBRanker(tree_method="hist", ndcg_exp_gain=True, n_estimators=2)
+    with pytest.raises(ValueError, match="Set `ndcg_exp_gain`"):
+        fit(ltr)
+    ltr = xgboost.XGBRanker(tree_method="hist", ndcg_exp_gain=False, n_estimators=2)
+    fit(ltr)
+
+
 def test_ranking_with_unweighted_data():
     Xrow = np.array([1, 2, 6, 8, 11, 14, 16, 17])
     Xcol = np.array([0, 0, 1, 1,  2,  2,  3,  3])