Optional normalization for learning to rank. (#10094)

doc/parameter.rst

@@ -500,7 +500,11 @@ These are parameters specific to learning to rank task. See :doc:`Learning to Ra
 
   It specifies the number of pairs sampled for each document when pair method is ``mean``, or the truncation level for queries when the pair method is ``topk``. For example, to train with ``ndcg@6``, set ``lambdarank_num_pair_per_sample`` to :math:`6` and ``lambdarank_pair_method`` to ``topk``.
 
-* ``lambdarank_unbiased`` [default = ``false``]
+* ``lambdarank_normalization`` [default = ``true``]
+
+  Whether to normalize the leaf value by lambda gradient. This can sometimes stagnate the training progress.
+
+*  ``lambdarank_unbiased`` [default = ``false``]
 
   Specify whether do we need to debias input click data.
 

doc/tutorials/learning_to_rank.rst

@@ -48,7 +48,7 @@ Notice that the samples are sorted based on their query index in a non-decreasin
   import xgboost as xgb
 
   # Make a synthetic ranking dataset for demonstration
-  seed = 1994 
+  seed = 1994
   X, y = make_classification(random_state=seed)
   rng = np.random.default_rng(seed)
   n_query_groups = 3
@@ -146,7 +146,8 @@ The consideration of effective pairs also applies to the choice of pair method (
 
 When using the mean strategy for generating pairs, where the target metric (like ``NDCG``) is computed over the whole query list, users can specify how many pairs should be generated per each document, by setting the ``lambdarank_num_pair_per_sample``. XGBoost will randomly sample ``lambdarank_num_pair_per_sample`` pairs for each element in the query group (:math:`|pairs| = |query| \times num\_pairsample`). Often, setting it to 1 can produce reasonable results. In cases where performance is inadequate due to insufficient number of effective pairs being generated, set ``lambdarank_num_pair_per_sample`` to a higher value. As more document pairs are generated, more effective pairs will be generated as well.
 
-On the other hand, if you are prioritizing the top :math:`k` documents, the ``lambdarank_num_pair_per_sample`` should be set slightly higher than :math:`k` (with a few more documents) to obtain a good training result.
+On the other hand, if you are prioritizing the top :math:`k` documents, the ``lambdarank_num_pair_per_sample`` should be set slightly higher than :math:`k` (with a few more documents) to obtain a good training result. Lastly, XGBoost employs additional regularization for learning to rank objectives, which can be disabled by setting the ``lambdarank_normalization`` to ``False``.
+
 
 **Summary** If you have large amount of training data:
 

python-package/xgboost/testing/ranking.py

@@ -100,3 +100,21 @@ def run_ranking_categorical(device: str) -> None:
     scores = cross_val_score(ltr, X, y)
     for s in scores:
         assert s > 0.7
+
+
+def run_normalization(device: str) -> None:
+    """Test normalization."""
+    X, y, qid, _ = tm.make_ltr(2048, 4, 64, 3)
+    ltr = xgb.XGBRanker(objective="rank:pairwise", n_estimators=4, device=device)
+    ltr.fit(X, y, qid=qid, eval_set=[(X, y)], eval_qid=[qid])
+    e0 = ltr.evals_result()
+
+    ltr = xgb.XGBRanker(
+        objective="rank:pairwise",
+        n_estimators=4,
+        device=device,
+        lambdarank_normalization=False,
+    )
+    ltr.fit(X, y, qid=qid, eval_set=[(X, y)], eval_qid=[qid])
+    e1 = ltr.evals_result()
+    assert e1["validation_0"]["ndcg@32"][-1] > e0["validation_0"]["ndcg@32"][-1]

src/common/ranking_utils.h

@@ -78,6 +78,7 @@ struct LambdaRankParam : public XGBoostParameter<LambdaRankParam> {
 
   // unbiased
   bool lambdarank_unbiased{false};
+  bool lambdarank_normalization{true};
   double lambdarank_bias_norm{1.0};
   // ndcg
   bool ndcg_exp_gain{true};
@@ -86,6 +87,7 @@ struct LambdaRankParam : public XGBoostParameter<LambdaRankParam> {
     return lambdarank_pair_method == that.lambdarank_pair_method &&
            lambdarank_num_pair_per_sample == that.lambdarank_num_pair_per_sample &&
            lambdarank_unbiased == that.lambdarank_unbiased &&
+           lambdarank_normalization == that.lambdarank_normalization &&
            lambdarank_bias_norm == that.lambdarank_bias_norm && ndcg_exp_gain == that.ndcg_exp_gain;
   }
   bool operator!=(LambdaRankParam const& that) const { return !(*this == that); }
@@ -134,6 +136,9 @@ struct LambdaRankParam : public XGBoostParameter<LambdaRankParam> {
     DMLC_DECLARE_FIELD(lambdarank_unbiased)
         .set_default(false)
         .describe("Unbiased lambda mart. Use extended IPW to debias click position");
+    DMLC_DECLARE_FIELD(lambdarank_normalization)
+        .set_default(true)
+        .describe("Whether to normalize the leaf value for lambda rank.");
     DMLC_DECLARE_FIELD(lambdarank_bias_norm)
         .set_default(1.0)
         .set_lower_bound(0.0)

src/objective/lambdarank_obj.cc

@@ -222,7 +222,7 @@ class LambdaRankObj : public FitIntercept {
     };
 
     MakePairs(ctx_, iter, p_cache_, g, g_label, g_rank, loop);
-    if (sum_lambda > 0.0) {
+    if (sum_lambda > 0.0 && param_.lambdarank_normalization) {
       double norm = std::log2(1.0 + sum_lambda) / sum_lambda;
       std::transform(g_gpair.Values().data(), g_gpair.Values().data() + g_gpair.Size(),
                      g_gpair.Values().data(), [norm](GradientPair const& g) { return g * norm; });

src/objective/lambdarank_obj.cu

@@ -266,12 +266,13 @@ void CalcGrad(Context const* ctx, MetaInfo const& info, std::shared_ptr<ltr::Ran
    */
   auto d_weights = common::MakeOptionalWeights(ctx, info.weights_);
   auto w_norm = p_cache->WeightNorm();
+  auto norm = p_cache->Param().lambdarank_normalization;
   thrust::for_each_n(ctx->CUDACtx()->CTP(), thrust::make_counting_iterator(0ul), d_gpair.Size(),
                      [=] XGBOOST_DEVICE(std::size_t i) mutable {
                        auto g = dh::SegmentId(d_gptr, i);
                        auto sum_lambda = thrust::get<2>(d_max_lambdas[g]);
                        // Normalization
-                       if (sum_lambda > 0.0) {
+                       if (sum_lambda > 0.0 && norm) {
                          double norm = std::log2(1.0 + sum_lambda) / sum_lambda;
                          d_gpair(i, 0) *= norm;
                        }

tests/python-gpu/test_gpu_ranking.py

@@ -6,6 +6,7 @@ import pytest
 
 import xgboost
 from xgboost import testing as tm
+from xgboost.testing.ranking import run_normalization
 
 pytestmark = tm.timeout(30)
 
@@ -126,3 +127,7 @@ def test_with_mq2008(objective, metric) -> None:
     dtest = xgboost.DMatrix(x_test, y_test, qid=qid_test)
 
     comp_training_with_rank_objective(dtrain, dtest, objective, metric)
+
+
+def test_normalization() -> None:
+    run_normalization("cuda")

tests/python/test_ranking.py

@@ -13,6 +13,7 @@ import xgboost
 from xgboost import testing as tm
 from xgboost.testing.data import RelDataCV, simulate_clicks, sort_ltr_samples
 from xgboost.testing.params import lambdarank_parameter_strategy
+from xgboost.testing.ranking import run_normalization
 
 
 def test_ndcg_custom_gain():
@@ -188,6 +189,10 @@ def test_unbiased() -> None:
     assert df["ti+"].iloc[-1] < df["ti+"].iloc[0]
 
 
+def test_normalization() -> None:
+    run_normalization("cpu")
+
+
 class TestRanking:
     @classmethod
     def setup_class(cls):