Parameter tests.

Author: trivialfis

python-package/xgboost/testing/params.py

@@ -47,3 +47,15 @@
         "max_cat_threshold": strategies.integers(1, 128),
     }
 )
+
+lambdarank_parameter_strategy = strategies.fixed_dictionaries(
+    {
+        "lambdarank_unbiased": strategies.sampled_from([True, False]),
+        "lambdarank_pair_method": strategies.sampled_from(["topk", "mean"]),
+        "lambdarank_num_pair_per_sample": strategies.integers(1, 8),
+        "lambdarank_bias_norm": strategies.floats(0.5, 2.0),
+        "objective": strategies.sampled_from(
+            ["rank:ndcg", "rank:map", "rank:pairwise"]
+        ),
+    }
+)

src/metric/rank_metric.cc

@@ -360,7 +360,10 @@ 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;
@@ -418,11 +421,6 @@ double Finalize(double score, double sw) {
 class EvalNDCG : public EvalRankWithCache<ltr::NDCGCache> {
  public:
   using EvalRankWithCache::EvalRankWithCache;
-
-  void Configure(Args const&) override {
-    // do not configure, otherwise the ndcg param will be forced into the same as the one in
-    // objective.
-  }
   const char* Name() const override { return name_.c_str(); }
 
   double Eval(HostDeviceVector<float> const& preds, MetaInfo const& info,
@@ -484,17 +482,6 @@ class EvalNDCG : public EvalRankWithCache<ltr::NDCGCache> {
 class EvalMAPScore : public EvalRankWithCache<ltr::MAPCache> {
  public:
   using EvalRankWithCache::EvalRankWithCache;
-
-  void Configure(Args const&) override {
-    // do not configure, otherwise the ndcg param will be forced into the same as the one in
-    // objective.
-  }
-  void SaveConfig(Json* p_out) const override {
-    auto& out = *p_out;
-    out["name"] = String{this->Name()};
-    out["map_param"] = ToJson(param_);
-  }
-  void LoadConfig(Json const& in) override { FromJson(in["map_param"], &param_); }
   const char* Name() const override { return name_.c_str(); }
 
   double Eval(HostDeviceVector<float> const& predt, MetaInfo const& info,

tests/python/test_ranking.py

@@ -5,9 +5,11 @@
 import numpy as np
 import pytest
 from scipy.sparse import csr_matrix
+from hypothesis import given, note, settings
 
 import xgboost
 from xgboost import testing as tm
+from xgboost.testing.params import lambdarank_parameter_strategy
 
 
 def test_ranking_with_unweighted_data():
@@ -74,6 +76,29 @@ def test_ranking_with_weighted_data():
         assert all(p <= q for p, q in zip(is_sorted, is_sorted[1:]))
 
 
+def test_error_msg(self) -> None:
+    X, y, qid, w = tm.make_ltr(10, 2, 2, 2)
+    ranker = xgboost.XGBRanker()
+    with pytest.raises(ValueError, match=r"equal to the number of query groups"):
+        ranker.fit(X, y, qid=qid, sample_weight=y)
+
+
+@given(lambdarank_parameter_strategy)
+@settings(deadline=None, print_blob=True)
+def test_lambdarank_parameters(params):
+    if params["objective"] == "rank:map":
+        rel = 2
+    else:
+        rel = 5
+    X, y, q, w = tm.make_ltr(4096, 3, 13, rel)
+    ranker = xgboost.XGBRanker(tree_method="hist", n_estimators=64, **params)
+    ranker.fit(X, y, qid=q, sample_weight=w, eval_set=[(X, y)], eval_qid=[q])
+    for k, v in ranker.evals_result()["validation_0"].items():
+        note(v)
+        assert v[-1] > v[0]
+        assert ranker.n_features_in_ == 3
+
+
 class TestRanking:
     @classmethod
     def setup_class(cls):
@@ -119,12 +144,6 @@ def teardown_class(cls):
         if os.path.exists(directory):
             shutil.rmtree(directory)
 
-    def test_error_msg(self) -> None:
-        X, y, qid, w = tm.make_ltr(10, 2, 2, 2)
-        ranker = xgboost.XGBRanker()
-        with pytest.raises(ValueError, match=r"equal to the number of query groups"):
-            ranker.fit(X, y, qid=qid, sample_weight=y)
-
     def test_training(self):
         """
         Train an XGBoost ranking model