Clear split info buffer in cost efficient gradient boosting before every iteration (fix partially #3679)

src/treelearner/cost_effective_gradient_boosting.hpp

@@ -10,6 +10,7 @@
 #include <LightGBM/dataset.h>
 #include <LightGBM/utils/common.h>
 #include <LightGBM/utils/log.h>
+#include <LightGBM/utils/threading.h>
 
 #include <vector>
 
@@ -32,6 +33,7 @@ class CostEfficientGradientBoosting {
       return true;
     }
   }
+
   void Init() {
     auto train_data = tree_learner_->train_data_;
     if (!init_) {
@@ -63,6 +65,17 @@ class CostEfficientGradientBoosting {
     }
     init_ = true;
   }
+
+  void BeforeTrain() {
+    // clear the splits in splits_per_leaf_
+    Threading::For<size_t>(0, splits_per_leaf_.size(), 1024,
+      [this] (int /*thread_index*/, size_t start, size_t end) {
+      for (size_t i = start; i < end; ++i) {
+        splits_per_leaf_[i].Reset();
+      }
+    });
+  }
+
   double DeltaGain(int feature_index, int real_fidx, int leaf_index,
                    int num_data_in_leaf, SplitInfo split_info) {
     auto config = tree_learner_->config_;
@@ -82,6 +95,7 @@ class CostEfficientGradientBoosting {
                      feature_index] = split_info;
     return delta;
   }
+
   void UpdateLeafBestSplits(Tree* tree, int best_leaf,
                             const SplitInfo* best_split_info,
                             std::vector<SplitInfo>* best_split_per_leaf) {

src/treelearner/serial_tree_learner.cpp

@@ -278,6 +278,10 @@ void SerialTreeLearner::BeforeTrain() {
   }
 
   larger_leaf_splits_->Init();
+
+  if (cegb_ != nullptr) {
+    cegb_->BeforeTrain();
+  }
 }
 
 bool SerialTreeLearner::BeforeFindBestSplit(const Tree* tree, int left_leaf, int right_leaf) {

tests/python_package_test/test_engine.py

@@ -3566,3 +3566,48 @@ def test_boost_from_average_with_single_leaf_trees():
     preds = model.predict(X)
     mean_preds = np.mean(preds)
     assert y.min() <= mean_preds <= y.max()
+
+
+def test_cegb_split_buffer_clean():
+    # modified from https://github.com/microsoft/LightGBM/issues/3679#issuecomment-938652811
+    # and https://github.com/microsoft/LightGBM/pull/5087
+    # test that the ``splits_per_leaf_`` of CEGB is cleaned before training a new tree
+    # which is done in the fix #5164
+    # without the fix:
+    #    Check failed: (best_split_info.left_count) > (0)
+
+    R, C = 1000, 100
+    seed = 29
+    np.random.seed(seed)
+    data = np.random.randn(R, C)
+    for i in range(1, C):
+        data[i] += data[0] * np.random.randn()
+
+    N = int(0.8 * len(data))
+    train_data = data[:N]
+    test_data = data[N:]
+    train_y = np.sum(train_data, axis=1)
+    test_y = np.sum(test_data, axis=1)
+
+    train = lgb.Dataset(train_data, train_y, free_raw_data=True)
+
+    params = {
+        'boosting_type': 'gbdt',
+        'objective': 'regression',
+        'max_bin': 255,
+        'num_leaves': 31,
+        'seed': 0,
+        'learning_rate': 0.1,
+        'min_data_in_leaf': 0,
+        'verbose': -1,
+        'min_split_gain': 1000.0,
+        'cegb_penalty_feature_coupled': 5 * np.arange(C),
+        'cegb_penalty_split': 0.0002,
+        'cegb_tradeoff': 10.0,
+        'force_col_wise': True,
+    }
+
+    model = lgb.train(params, train, num_boost_round=10)
+    predicts = model.predict(test_data)
+    rmse = np.sqrt(mean_squared_error(test_y, predicts))
+    assert rmse < 10.0