remove.

src/objective/rank_obj.cc

@@ -183,6 +183,7 @@ class LambdaMARTNDCG : public FitIntercept {
       auto gpairs = h_gpair.subspan(info.group_ptr_[g], cnt);
       auto labels = h_label.Slice(make_range(g), 0);
       if (ndcg_param_.lambdamart_unbiased) {
+        CHECK(false);
         this->CalcLambdaForGroup<true>(predts, labels, gpairs, info, g);
       } else {
         this->CalcLambdaForGroup<false>(predts, labels, gpairs, info, g);

src/objective/rank_obj.h

@@ -75,29 +75,10 @@ XGBOOST_DEVICE inline void LambdaNDCG(linalg::TensorView<float const, 1> labels,
   // fixme: cache discounts
   double delta_NDCG = std::abs(DeltaNDCG(y_high, y_low, rank_high, rank_low, inv_IDCG));
 
-  if (unbiased) {
-    // fixme: is this necessary?
-    delta_NDCG /= std::max(score_distance, 0.01);
-
-    double cost = std::log(1.0 / (1.0 - sigmoid)) * delta_NDCG;
-
-    int debias_high_rank = static_cast<int>(std::min(rank_high, p.lambdamart_truncation - 1));
-    int debias_low_rank = static_cast<int>(std::min(rank_low, p.lambdamart_truncation - 1));
-
-    li(debias_high_rank) += cost / tj(debias_low_rank);
-    lj(debias_low_rank) += cost / ti(debias_high_rank);
-  }
-
   constexpr double kEps = 1e-16f;
   auto lambda_ij = (sigmoid - 1.0) * delta_NDCG;
   auto hessian_ij = std::max(sigmoid * (1.0 - sigmoid), kEps) * delta_NDCG;
 
-  if (unbiased) {
-    lambda_ij /= (tj(rank_high) * ti(rank_low));
-    hessian_ij /= (tj(rank_high) * ti(rank_low));
-  }
-
-  std::cout << "i:" << i << "j:" << j << ", h:" << idx_high << " l:" << idx_low << ", |Z|:" << delta_NDCG << ", lij:" << lambda_ij << ", hij:" << hessian_ij << std::endl;
   auto pg = GradientPair{static_cast<float>(lambda_ij), static_cast<float>(hessian_ij)};
   auto ng = GradientPair{static_cast<float>(-lambda_ij), static_cast<float>(hessian_ij)};