Fix init with scale pos weight. (dmlc#11280)

Author: trivialfis

src/objective/init_estimation.h

@@ -13,7 +13,7 @@ class FitIntercept : public ObjFunction {
   void InitEstimation(MetaInfo const& info, linalg::Vector<float>* base_score) const override;
 };
 
-class FitInterceptGlmLike : public ObjFunction {
+class FitInterceptGlmLike : public FitIntercept {
  public:
   void InitEstimation(MetaInfo const& info, linalg::Vector<float>* base_score) const override;
 };

src/objective/regression_obj.cu

@@ -187,6 +187,16 @@ class RegLossObj : public FitInterceptGlmLike {
         .Eval(io_preds);
   }
 
+  void InitEstimation(MetaInfo const& info, linalg::Vector<float>* base_score) const override {
+    if (std::abs(this->param_.scale_pos_weight - 1.0f) > kRtEps) {
+      // Use newton method if `scale_pos_weight` is present. The alternative is to use
+      // weighted mean, but we also need to take sample weight into account.
+      FitIntercept::InitEstimation(info, base_score);
+    } else {
+      FitInterceptGlmLike::InitEstimation(info, base_score);
+    }
+  }
+
   [[nodiscard]] float ProbToMargin(float base_score) const override {
     return Loss::ProbToMargin(base_score);
   }

tests/python/test_objectives.py

@@ -18,3 +18,12 @@ def test_exp_family() -> None:
     )
     # The base score stored in the booster model is un-transformed
     np.testing.assert_allclose([get_basescore(m) for m in (reg, clf, clf1)], y.mean())
+
+    X, y = make_classification(weights=[0.8, 0.2], random_state=2025)
+    clf = xgb.train(
+        {"objective": "binary:logistic", "scale_pos_weight": 4.0},
+        xgb.QuantileDMatrix(X, y),
+        num_boost_round=1,
+    )
+    score = get_basescore(clf)
+    np.testing.assert_allclose(score, 0.5, rtol=1e-3)