fix bug in finding best split for categorical feature.

src/treelearner/feature_histogram.hpp

@@ -131,12 +131,12 @@ public:
       output->left_output = CalculateSplittedLeafOutput(best_sum_left_gradient, best_sum_left_hessian);
       output->left_count = best_left_count;
       output->left_sum_gradient = best_sum_left_gradient;
-      output->left_sum_hessian = best_sum_left_hessian;
+      output->left_sum_hessian = best_sum_left_hessian - kEpsilon;
       output->right_output = CalculateSplittedLeafOutput(sum_gradient - best_sum_left_gradient,
         sum_hessian - best_sum_left_hessian);
       output->right_count = num_data - best_left_count;
       output->right_sum_gradient = sum_gradient - best_sum_left_gradient;
-      output->right_sum_hessian = sum_hessian - best_sum_left_hessian;
+      output->right_sum_hessian = sum_hessian - best_sum_left_hessian - kEpsilon;
       output->gain = best_gain - gain_shift;
     } else {
       output->feature = meta_->feature_idx;
@@ -148,6 +148,9 @@ public:
     SplitInfo* output) {
     double best_gain = kMinScore;
     uint32_t best_threshold = static_cast<uint32_t>(meta_->num_bin);
+    data_size_t best_left_count = 0;
+    double best_sum_left_gradient = 0.0f;
+    double best_sum_left_hessian = 0.0f;
     double gain_shift = GetLeafSplitGain(sum_gradient, sum_hessian);
     double min_gain_shift = gain_shift + meta_->tree_config->min_gain_to_split;
     is_splittable_ = false;
@@ -179,6 +182,9 @@ public:
       // better split point
       if (current_gain > best_gain) {
         best_threshold = static_cast<uint32_t>(t + bias);
+        best_sum_left_gradient = data_[t].sum_gradients;
+        best_sum_left_hessian = data_[t].sum_hessians + kEpsilon;
+        best_left_count = data_[t].cnt;
         best_gain = current_gain;
       }
     }
@@ -186,7 +192,7 @@ public:
     if (bias == 1) {
       t = meta_->num_bin - 1 - bias;
       double sum_bin0_gradient = sum_gradient;
-      double sum_bin0_hessian = sum_hessian;
+      double sum_bin0_hessian = sum_hessian - 2 * kEpsilon;
       data_size_t cnt_bin0 = num_data;
       for (; t >= 0; --t) {
         sum_bin0_gradient -= data_[t].sum_gradients;
@@ -207,6 +213,9 @@ public:
           // better split point
           if (current_gain > best_gain) {
             best_threshold = static_cast<uint32_t>(0);
+            best_sum_left_gradient = sum_bin0_gradient;
+            best_sum_left_hessian = sum_bin0_hessian + kEpsilon;
+            best_left_count = cnt_bin0;
             best_gain = current_gain;
           }
         }
@@ -216,17 +225,15 @@ public:
       // update split information
       output->feature = meta_->feature_idx;
       output->threshold = best_threshold;
-      output->left_output = CalculateSplittedLeafOutput(data_[best_threshold].sum_gradients,
-        data_[best_threshold].sum_hessians + kEpsilon);
-      output->left_count = data_[best_threshold].cnt;
-      output->left_sum_gradient = data_[best_threshold].sum_gradients;
-      output->left_sum_hessian = data_[best_threshold].sum_hessians + kEpsilon;
-
-      output->right_output = CalculateSplittedLeafOutput(sum_gradient - data_[best_threshold].sum_gradients,
-        sum_hessian - data_[best_threshold].sum_hessians - kEpsilon);
-      output->right_count = num_data - data_[best_threshold].cnt;
-      output->right_sum_gradient = sum_gradient - data_[best_threshold].sum_gradients;
-      output->right_sum_hessian = sum_hessian - data_[best_threshold].sum_hessians - kEpsilon;
+      output->left_output = CalculateSplittedLeafOutput(best_sum_left_gradient, best_sum_left_hessian);
+      output->left_count = best_left_count;
+      output->left_sum_gradient = best_sum_left_gradient;
+      output->left_sum_hessian = best_sum_left_hessian - kEpsilon;
+      output->right_output = CalculateSplittedLeafOutput(sum_gradient - best_sum_left_gradient,
+        sum_hessian - best_sum_left_hessian);
+      output->right_count = num_data - best_left_count;
+      output->right_sum_gradient = sum_gradient - best_sum_left_gradient;
+      output->right_sum_hessian = sum_hessian - best_sum_left_hessian - kEpsilon;
       output->gain = best_gain - gain_shift;
     } else {
       output->feature = meta_->feature_idx;