enforce interaction constraints with monotone_constraints_method = intermediate/advanced (#4043)

* add test for interaction constraints and monotone constraints

* enforce interaction constraints in RecomputeBestSplitForLeaf

* code formatting

* code formatting

* move interaction constraint test to test_engine

* Apply suggestions from code review
Co-authored-by: Nikita Titov <nekit94-08@mail.ru>
Co-authored-by: Nikita Titov <nekit94-08@mail.ru>

src/treelearner/serial_tree_learner.cpp

@@ -677,7 +677,7 @@ void SerialTreeLearner::SplitInner(Tree* tree, int best_leaf, int* left_leaf,
       best_split_per_leaf_);
   // update leave outputs if needed
   for (auto leaf : leaves_need_update) {
-    RecomputeBestSplitForLeaf(leaf, &best_split_per_leaf_[leaf]);
+    RecomputeBestSplitForLeaf(tree, leaf, &best_split_per_leaf_[leaf]);
   }
 }
 
@@ -768,7 +768,7 @@ double SerialTreeLearner::GetParentOutput(const Tree* tree, const LeafSplits* le
   return parent_output;
 }
 
-void SerialTreeLearner::RecomputeBestSplitForLeaf(int leaf, SplitInfo* split) {
+void SerialTreeLearner::RecomputeBestSplitForLeaf(Tree* tree, int leaf, SplitInfo* split) {
   FeatureHistogram* histogram_array_;
   if (!histogram_pool_.Get(leaf, &histogram_array_)) {
     Log::Warning(
@@ -795,6 +795,7 @@ void SerialTreeLearner::RecomputeBestSplitForLeaf(int leaf, SplitInfo* split) {
 
   OMP_INIT_EX();
 // find splits
+std::vector<int8_t> node_used_features = col_sampler_.GetByNode(tree, leaf);
 #pragma omp parallel for schedule(static) num_threads(share_state_->num_threads)
   for (int feature_index = 0; feature_index < num_features_; ++feature_index) {
     OMP_LOOP_EX_BEGIN();
@@ -804,7 +805,7 @@ void SerialTreeLearner::RecomputeBestSplitForLeaf(int leaf, SplitInfo* split) {
     }
     const int tid = omp_get_thread_num();
     int real_fidx = train_data_->RealFeatureIndex(feature_index);
-    ComputeBestSplitForFeature(histogram_array_, feature_index, real_fidx, true,
+    ComputeBestSplitForFeature(histogram_array_, feature_index, real_fidx, node_used_features[feature_index],
                                num_data, &leaf_splits, &bests[tid], parent_output);
 
     OMP_LOOP_EX_END();

src/treelearner/serial_tree_learner.h

@@ -128,7 +128,7 @@ class SerialTreeLearner: public TreeLearner {
 
   void GetShareStates(const Dataset* dataset, bool is_constant_hessian, bool is_first_time);
 
-  void RecomputeBestSplitForLeaf(int leaf, SplitInfo* split);
+  void RecomputeBestSplitForLeaf(Tree* tree, int leaf, SplitInfo* split);
 
   /*!
   * \brief Some initial works before training

tests/python_package_test/test_engine.py

@@ -1252,7 +1252,8 @@ def generate_trainset_for_monotone_constraints_tests(x3_to_category=True):
     return trainset
 
 
-def test_monotone_constraints():
+@pytest.mark.parametrize("test_with_interaction_constraints", [True, False])
+def test_monotone_constraints(test_with_interaction_constraints):
     def is_increasing(y):
         return (np.diff(y) >= 0.0).all()
 
@@ -1273,28 +1274,69 @@ def test_monotone_constraints():
             monotonically_increasing_y = learner.predict(monotonically_increasing_x)
             monotonically_decreasing_x = np.column_stack((fixed_x, variable_x, fixed_x))
             monotonically_decreasing_y = learner.predict(monotonically_decreasing_x)
-            non_monotone_x = np.column_stack((fixed_x,
+            non_monotone_x = np.column_stack(
+                (
                     fixed_x,
-                                              categorize(variable_x) if x3_to_category else variable_x))
+                    fixed_x,
+                    categorize(variable_x) if x3_to_category else variable_x,
+                )
+            )
             non_monotone_y = learner.predict(non_monotone_x)
-            if not (is_increasing(monotonically_increasing_y)
+            if not (
+                is_increasing(monotonically_increasing_y)
                 and is_decreasing(monotonically_decreasing_y)
-                    and is_non_monotone(non_monotone_y)):
+                and is_non_monotone(non_monotone_y)
+            ):
                 return False
         return True
 
+    def are_interactions_enforced(gbm, feature_sets):
+        def parse_tree_features(gbm):
+            # trees start at position 1.
+            tree_str = gbm.model_to_string().split("Tree")[1:]
+            feature_sets = []
+            for tree in tree_str:
+                # split_features are in 4th line.
+                features = tree.splitlines()[3].split("=")[1].split(" ")
+                features = set(f"Column_{f}" for f in features)
+                feature_sets.append(features)
+            return np.array(feature_sets)
+
+        def has_interaction(treef):
+            n = 0
+            for fs in feature_sets:
+                if len(treef.intersection(fs)) > 0:
+                    n += 1
+            return n > 1
+
+        tree_features = parse_tree_features(gbm)
+        has_interaction_flag = np.array(
+            [has_interaction(treef) for treef in tree_features]
+        )
+
+        return not has_interaction_flag.any()
+
     for test_with_categorical_variable in [True, False]:
-        trainset = generate_trainset_for_monotone_constraints_tests(test_with_categorical_variable)
+        trainset = generate_trainset_for_monotone_constraints_tests(
+            test_with_categorical_variable
+        )
         for monotone_constraints_method in ["basic", "intermediate", "advanced"]:
             params = {
-                'min_data': 20,
-                'num_leaves': 20,
-                'monotone_constraints': [1, -1, 0],
+                "min_data": 20,
+                "num_leaves": 20,
+                "monotone_constraints": [1, -1, 0],
                 "monotone_constraints_method": monotone_constraints_method,
                 "use_missing": False,
             }
+            if test_with_interaction_constraints:
+                params["interaction_constraints"] = [[0], [1], [2]]
             constrained_model = lgb.train(params, trainset)
-            assert is_correctly_constrained(constrained_model, test_with_categorical_variable)
+            assert is_correctly_constrained(
+                constrained_model, test_with_categorical_variable
+            )
+            if test_with_interaction_constraints:
+                feature_sets = [["Column_0"], ["Column_1"], "Column_2"]
+                assert are_interactions_enforced(constrained_model, feature_sets)
 
 
 def test_monotone_penalty():