[BP] Fix unbiased ltr with training continuation. (dmlc#10908) (dmlc#10909)

Author: trivialfis

src/objective/lambdarank_obj.cc

@@ -314,7 +314,7 @@ class LambdaRankObj : public FitIntercept {
       CHECK_EQ(info.weights_.Size(), n_groups) << error::GroupWeight();
     }
 
-    if (ti_plus_.Size() == 0 && param_.lambdarank_unbiased) {
+    if ((ti_plus_.Empty() || li_full_.Empty()) && param_.lambdarank_unbiased) {
       CHECK_EQ(iter, 0);
       ti_plus_ = linalg::Constant<double>(ctx_, 1.0, p_cache_->MaxPositionSize());
       tj_minus_ = linalg::Constant<double>(ctx_, 1.0, p_cache_->MaxPositionSize());

tests/python/test_ranking.py

@@ -202,6 +202,12 @@ def after_training(self, model) -> bool:
     # less biased on low ranks.
     assert df["ti+"].iloc[-1] < df["ti+"].iloc[0]
 
+    # Training continuation
+    ltr.fit(x, c, qid=q, eval_set=[(x, c)], eval_qid=[q], xgb_model=ltr)
+    # normalized
+    np.testing.assert_allclose(df["ti+"].iloc[0], 1.0)
+    np.testing.assert_allclose(df["tj-"].iloc[0], 1.0)
+
 
 def test_normalization() -> None:
     run_normalization("cpu")