change default missing handling to "separate"

interpretml · Dec 27, 2024 · 8581df3 · 8581df3
1 parent 4204754
commit 8581df3
Show file tree

Hide file tree

Showing 2 changed files with 47 additions and 48 deletions.
diff --git a/python/interpret-core/interpret/glassbox/_ebm/_ebm.py b/python/interpret-core/interpret/glassbox/_ebm/_ebm.py
@@ -2771,7 +2771,7 @@ class ExplainableBoostingClassifier(ClassifierMixin, EBMModel):
         L2 regularization.
     max_delta_step : float, default=0.0
         Used to limit the max output of tree leaves. <=0.0 means no constraint.
-    missing: str, default="low"
+    missing: str, default="separate"
 
         Method for handling missing values during boosting. The placement of the missing value bin can influence
         the resulting model graphs. For example, placing the bin on the "low" side may cause missing values to
@@ -2944,7 +2944,7 @@ def __init__(
         reg_alpha: Optional[float] = 0.0,
         reg_lambda: Optional[float] = 0.0,
         max_delta_step: Optional[float] = 0.0,
-        missing: str = "low",
+        missing: str = "separate",
         max_leaves: int = 3,
         monotone_constraints: Optional[Sequence[int]] = None,
         objective: str = "log_loss",
@@ -3143,7 +3143,7 @@ class ExplainableBoostingRegressor(RegressorMixin, EBMModel):
         L2 regularization.
     max_delta_step : float, default=0.0
         Used to limit the max output of tree leaves. <=0.0 means no constraint.
-    missing: str, default="low"
+    missing: str, default="separate"
 
         Method for handling missing values during boosting. The placement of the missing value bin can influence
         the resulting model graphs. For example, placing the bin on the "low" side may cause missing values to
@@ -3316,7 +3316,7 @@ def __init__(
         reg_alpha: Optional[float] = 0.0,
         reg_lambda: Optional[float] = 0.0,
         max_delta_step: Optional[float] = 0.0,
-        missing: str = "low",
+        missing: str = "separate",
         max_leaves: int = 2,
         monotone_constraints: Optional[Sequence[int]] = None,
         objective: str = "rmse",

diff --git a/shared/libebm/PartitionOneDimensionalBoosting.cpp b/shared/libebm/PartitionOneDimensionalBoosting.cpp
@@ -326,50 +326,7 @@ static ErrorEbm Flatten(BoosterShell* const pBoosterShell,
 
          while(true) {
             if(nullptr == pTreeNode) {
-            done:;
-               EBM_ASSERT(cSamplesTotalDebug == cSamplesExpectedDebug);
-
-               EBM_ASSERT(bNominal || pUpdateScore == aUpdateScore + cScores * cSlices);
-
-               EBM_ASSERT(bNominal || pSplit == cSlices - 1 + pInnerTermUpdate->GetSplitPointer(iDimension));
-
-#ifndef NDEBUG
-               UIntSplit prevDebug = 0;
-               for(size_t iDebug = 0; iDebug < cSlices - 1; ++iDebug) {
-                  UIntSplit curDebug = pInnerTermUpdate->GetSplitPointer(iDimension)[iDebug];
-                  EBM_ASSERT(prevDebug < curDebug);
-                  prevDebug = curDebug;
-               }
-               EBM_ASSERT(prevDebug < cBins);
-#endif
-
-               EBM_ASSERT(nullptr == pMissingValueTreeNode || nullptr != pMissingBin);
-               if(nullptr != pMissingBin) {
-                  EBM_ASSERT(bMissing);
-
-                  FloatScore hess = static_cast<FloatCalc>(pMissingBin->GetWeight());
-                  const auto* pGradientPair = pMissingBin->GetGradientPairs();
-                  const auto* const pGradientPairEnd = pGradientPair + cScores;
-                  FloatScore* pMissingUpdateScore = aUpdateScore;
-                  do {
-                     if(bUpdateWithHessian) {
-                        hess = static_cast<FloatCalc>(pGradientPair->GetHess());
-                     }
-                     FloatCalc updateScore = -CalcNegUpdate<true>(static_cast<FloatCalc>(pGradientPair->m_sumGradients),
-                           hess,
-                           regAlpha,
-                           regLambda,
-                           deltaStepMax);
-
-                     *pMissingUpdateScore = updateScore;
-                     ++pMissingUpdateScore;
-
-                     ++pGradientPair;
-                  } while(pGradientPairEnd != pGradientPair);
-               }
-
-               LOG_0(Trace_Verbose, "Exited Flatten");
-               return Error_None;
+               goto done;
             }
             if(!pTreeNode->DECONSTRUCT_IsRightChildTraversal()) {
                // we checked earlier that countBins could be converted to a UIntSplit
@@ -411,6 +368,48 @@ static ErrorEbm Flatten(BoosterShell* const pBoosterShell,
          }
       }
    }
+
+done:;
+   EBM_ASSERT(cSamplesTotalDebug == cSamplesExpectedDebug);
+
+   EBM_ASSERT(bNominal || pUpdateScore == aUpdateScore + cScores * cSlices);
+
+   EBM_ASSERT(bNominal || pSplit == cSlices - 1 + pInnerTermUpdate->GetSplitPointer(iDimension));
+
+#ifndef NDEBUG
+   UIntSplit prevDebug = 0;
+   for(size_t iDebug = 0; iDebug < cSlices - 1; ++iDebug) {
+      UIntSplit curDebug = pInnerTermUpdate->GetSplitPointer(iDimension)[iDebug];
+      EBM_ASSERT(prevDebug < curDebug);
+      prevDebug = curDebug;
+   }
+   EBM_ASSERT(prevDebug < cBins);
+#endif
+
+   EBM_ASSERT(nullptr == pMissingValueTreeNode || nullptr != pMissingBin);
+   if(nullptr != pMissingBin) {
+      EBM_ASSERT(bMissing);
+
+      FloatScore hess = static_cast<FloatCalc>(pMissingBin->GetWeight());
+      const auto* pGradientPair = pMissingBin->GetGradientPairs();
+      const auto* const pGradientPairEnd = pGradientPair + cScores;
+      FloatScore* pMissingUpdateScore = aUpdateScore;
+      do {
+         if(bUpdateWithHessian) {
+            hess = static_cast<FloatCalc>(pGradientPair->GetHess());
+         }
+         FloatCalc updateScore = -CalcNegUpdate<true>(
+               static_cast<FloatCalc>(pGradientPair->m_sumGradients), hess, regAlpha, regLambda, deltaStepMax);
+
+         *pMissingUpdateScore = updateScore;
+         ++pMissingUpdateScore;
+
+         ++pGradientPair;
+      } while(pGradientPairEnd != pGradientPair);
+   }
+
+   LOG_0(Trace_Verbose, "Exited Flatten");
+   return Error_None;
 }
 WARNING_POP