[YDF] Deprecate SparseObliqueSplit.binary_weights hyperparameter

In favor of `SparseObliqueSplit.weights`.

PiperOrigin-RevId: 701854322

CHANGELOG.md

@@ -15,6 +15,8 @@ Changelog under `yggdrasil_decision_forests/port/python/CHANGELOG.md`.
 -   Add support for distributed training for ranking gradient boosted tree
     models.
 -   Add support for AVRO data file using the "avro:" prefix.
+-   Deprecated `SparseObliqueSplit.binary_weights` hyperparameter in favor of
+    `SparseObliqueSplit.weights`.
 
 ### Misc
 

yggdrasil_decision_forests/learner/decision_tree/decision_tree.proto

@@ -220,10 +220,28 @@ message DecisionTreeTrainingConfig {
     // This value is best tuned for each dataset.
     optional float projection_density_factor = 3 [default = 2];
 
+    // Deprecated, use `weights` instead.
+    //
     // If true, the weight will be sampled in {-1,1} (default in  "Sparse
     // Projection Oblique Random Forests" (Tomita et al, 2020)). If false, the
     // weight will be sampled in [-1,1].
-    optional bool binary_weight = 4 [default = true];
+    optional bool binary_weight = 4 [default = true, deprecated = true];
+
+    // Weights to apply to the projections.
+    //
+    // Continuous weights generally give better performance.
+    oneof weights {
+      BinaryWeights binary = 7;
+      ContinuousWeights continuous = 8;
+    }
+
+    // Weights sample in {-1, 1} (default in "Sparse Projection Oblique Random
+    // Forests" (Tomita et al, 2020))).
+    message BinaryWeights {}
+
+    // Weights sample in [-1, 1]. Consistently gives better quality models than
+    // binary weights.
+    message ContinuousWeights {}
 
     // Normalization applied on the features, before applying the sparse oblique
     // projections.

yggdrasil_decision_forests/learner/decision_tree/generic_parameters.cc

@@ -308,9 +308,7 @@ Increasing this value increases training and inference time (on average). This v
     ASSIGN_OR_RETURN(auto param,
                      get_params(kHParamSplitAxisSparseObliqueWeights));
     param->mutable_categorical()->set_default_value(
-        config.sparse_oblique_split().binary_weight()
-            ? kHParamSplitAxisSparseObliqueWeightsBinary
-            : kHParamSplitAxisSparseObliqueWeightsContinuous);
+        kHParamSplitAxisSparseObliqueWeightsBinary);
     param->mutable_categorical()->add_possible_values(
         kHParamSplitAxisSparseObliqueWeightsBinary);
     param->mutable_categorical()->add_possible_values(
@@ -744,9 +742,9 @@ absl::Status SetHyperParameters(
       if (dt_config->has_sparse_oblique_split()) {
         const auto& value = hparam.value().value().categorical();
         if (value == kHParamSplitAxisSparseObliqueWeightsBinary) {
-          dt_config->mutable_sparse_oblique_split()->set_binary_weight(true);
+          dt_config->mutable_sparse_oblique_split()->mutable_binary();
         } else if (value == kHParamSplitAxisSparseObliqueWeightsContinuous) {
-          dt_config->mutable_sparse_oblique_split()->set_binary_weight(false);
+          dt_config->mutable_sparse_oblique_split()->mutable_continuous();
         } else {
           return absl::InvalidArgumentError(absl::StrCat(
               "Unknown value for parameter ",

yggdrasil_decision_forests/learner/decision_tree/oblique.cc

@@ -676,7 +676,9 @@ void SampleProjection(const absl::Span<const int>& features,
 
   const auto gen_weight = [&](const int feature) -> float {
     float weight = unif1m1(*random);
-    if (dt_config.sparse_oblique_split().binary_weight()) {
+    if (dt_config.sparse_oblique_split().has_binary() ||
+        dt_config.sparse_oblique_split().weights_case() ==
+            dt_config.sparse_oblique_split().WEIGHTS_NOT_SET) {
       weight = (weight >= 0) ? 1.f : -1.f;
     }
 

yggdrasil_decision_forests/learner/decision_tree/training.cc

@@ -3836,6 +3836,15 @@ void SetDefaultHyperParameters(proto::DecisionTreeTrainingConfig* config) {
   }
 
   config->mutable_internal()->set_sorting_strategy(sorting_strategy);
+
+  if (config->sparse_oblique_split().has_binary_weight()) {
+    if (config->sparse_oblique_split().binary_weight()) {
+      config->mutable_sparse_oblique_split()->mutable_binary();
+    } else {
+      config->mutable_sparse_oblique_split()->mutable_continuous();
+    }
+    config->mutable_sparse_oblique_split()->clear_binary_weight();
+  }
 }
 
 template <class T, class S, class C>
@@ -4064,6 +4073,17 @@ absl::Status DecisionTreeTrain(
         "pure_serving_model=true.");
   }
 
+  // Check if oblique splits are correctly specified
+  if (dt_config.sparse_oblique_split().has_binary_weight() &&
+      dt_config.sparse_oblique_split().weights_case() !=
+          dt_config.sparse_oblique_split().WEIGHTS_NOT_SET) {
+    return absl::InvalidArgumentError(
+        "Both sparse_oblique_split.binary_weights and "
+        "sparse_oblique_split.weights are set. Setting "
+        "sparse_oblique_split.binary_weights is deprecated and replaced by "
+        "just setting sparse_oblique_split.weights.");
+  }
+
   if (dt_config.has_honest()) {
     // Split the examples in two parts. One ("selected_examples_buffer") will be
     // used to infer the structure of the trees while the second

yggdrasil_decision_forests/port/python/ydf/learner/learner_test.py

@@ -1052,6 +1052,50 @@ def test_tuner_predefined(self):
     self.assertIsNotNone(logs)
     self.assertLen(logs.trials, 5)
 
+  def test_oblique_weights_default(self):
+    learner = specialized_learners.CartLearner(
+        label="label",
+        max_depth=2,
+        split_axis="SPARSE_OBLIQUE",
+    )
+    f1 = np.linspace(-1, 1, 50) ** 2
+    f2 = np.linspace(1.5, -0.5, 50) ** 2
+    label = (0.2 * f1 + 0.7 * f2 >= 0.25).astype(int)
+    ds = {"f1": f1, "f2": f2, "label": label}
+    model = learner.train(ds)
+    root_weights = model.get_tree(0).root.condition.weights
+    self.assertTrue(all(x in (-1.0, 1.0) for x in root_weights))
+
+  def test_oblique_weights_binary(self):
+    learner = specialized_learners.CartLearner(
+        label="label",
+        max_depth=2,
+        split_axis="SPARSE_OBLIQUE",
+        sparse_oblique_weights="BINARY",
+    )
+    f1 = np.linspace(-1, 1, 50) ** 2
+    f2 = np.linspace(1.5, -0.5, 50) ** 2
+    label = (0.2 * f1 + 0.7 * f2 >= 0.25).astype(int)
+    ds = {"f1": f1, "f2": f2, "label": label}
+    model = learner.train(ds)
+    root_weights = model.get_tree(0).root.condition.weights
+    self.assertTrue(all(x in (-1.0, 1.0) for x in root_weights))
+
+  def test_oblique_weights_continuous(self):
+    learner = specialized_learners.CartLearner(
+        label="label",
+        max_depth=2,
+        split_axis="SPARSE_OBLIQUE",
+        sparse_oblique_weights="CONTINUOUS",
+    )
+    f1 = np.linspace(-1, 1, 50) ** 2
+    f2 = np.linspace(1.5, -0.5, 50) ** 2
+    label = (0.2 * f1 + 0.7 * f2 >= 0.25).astype(int)
+    ds = {"f1": f1, "f2": f2, "label": label}
+    model = learner.train(ds)
+    root_weights = model.get_tree(0).root.condition.weights
+    self.assertFalse(all(x in (-1.0, 1.0) for x in root_weights))
+
 
 class GradientBoostedTreesLearnerTest(LearnerTest):