Почистил примеры

InstallerSamples/MachineLearning/08_Datasets/MoscowHousing/06_RandomForest.pas

@@ -7,7 +7,7 @@ begin
   var features := ['rooms','area','kitchen_area','floor','floors_total','metro_minutes'];
   var target := 'price';
 
-  var (trainDf, testDf) := df.TrainTestSplit(0.2, 42);
+  var (trainDf, testDf) := df.TrainTestSplit(0.2, seed := 42);
 
   var Xtrain := trainDf.ToMatrix(features);
   var ytrain := trainDf.ToVector(target);

InstallerSamples/MachineLearning/08_Datasets/MoscowHousing/07_GradientBoosting.pas

@@ -7,7 +7,7 @@ begin
   var features := ['rooms','area','kitchen_area','floor','floors_total','metro_minutes'];
   var target := 'price';
 
-  var (trainDf, testDf) := df.TrainTestSplit(0.2, 42);
+  var (trainDf, testDf) := df.TrainTestSplit(0.2, seed := 42);
 
   var Xtrain := trainDf.ToMatrix(features);
   var ytrain := trainDf.ToVector(target);

InstallerSamples/MachineLearning/08_Datasets/MoscowHousing/10_FeatureImportance1.pas

@@ -11,9 +11,10 @@ begin
 
   var pipe :=
     DataPipeline.Build(
+      TaskKind.tkRegression,
       target,
       features,
-      new LabelEncoder('renovation'),
+      new OrdinalEncoder('renovation'),
       model
     );
 

InstallerSamples/MachineLearning/08_Datasets/TitanicRu/06_ModelComparison.pas

@@ -12,10 +12,10 @@ begin
   df := portImputer.FitTransform(df);
 
   // Кодируем категориальные признаки числами.
-  var sexEncoder := new LabelEncoder('Пол');
+  var sexEncoder := new OrdinalEncoder('Пол');
   df := sexEncoder.FitTransform(df);
 
-  var portEncoder := new LabelEncoder('ПортПосадки');
+  var portEncoder := new OrdinalEncoder('ПортПосадки');
   df := portEncoder.FitTransform(df);
 
   var features := ['Класс', 'Пол', 'Возраст', 'БратьяИСупруги', 'РодителиИДети', 'ЦенаБилета', 'ПортПосадки'];

InstallerSamples/MachineLearning/08_Datasets/TitanicRu/09_FeatureImportance.pas

@@ -10,10 +10,10 @@ begin
   var portImputer := new Imputer('Саутгемптон', ['ПортПосадки']);
   df := portImputer.FitTransform(df);
 
-  var sexEncoder := new LabelEncoder('Пол');
+  var sexEncoder := new OrdinalEncoder('Пол');
   df := sexEncoder.FitTransform(df);
 
-  var portEncoder := new LabelEncoder('ПортПосадки');
+  var portEncoder := new OrdinalEncoder('ПортПосадки');
   df := portEncoder.FitTransform(df);
 
   var features := ['Класс', 'Пол', 'Возраст', 'БратьяИСупруги', 'РодителиИДети', 'ЦенаБилета', 'ПортПосадки'];

InstallerSamples/MachineLearning/13_Seminars/seminar1.pas

@@ -11,10 +11,10 @@ begin
   df := imputer.FitTransform(df);
   
   // Кодируем категориальные признаки
-  var encoder := new LabelEncoder('region_name');
+  var encoder := new OrdinalEncoder('region_name');
   df := encoder.FitTransform(df);
   
-  var encoder2 := new LabelEncoder('federal_district');
+  var encoder2 := new OrdinalEncoder('federal_district');
   df := encoder2.FitTransform(df);
   
   df.Print;

InstallerSamples/MachineLearning/13_Seminars/seminar2.pas

@@ -16,10 +16,10 @@ begin
   var imputer := new Imputer(['population', 'lat', 'lon']);
   df := imputer.FitTransform(df);
   
-  var le1 := new LabelEncoder('region_name');
+  var le1 := new OrdinalEncoder('region_name');
   df := le1.FitTransform(df);
   
-  var le2 := new LabelEncoder('federal_district');
+  var le2 := new OrdinalEncoder('federal_district');
   df := le2.FitTransform(df);
   
   // Явно задаём признаки

InstallerSamples/MachineLearning/13_Seminars/seminar3.pas

@@ -7,10 +7,10 @@ begin
   var imputer := new Imputer(['population', 'lat', 'lon']);
   df := imputer.FitTransform(df);
   
-  var le1 := new LabelEncoder('region_name');
+  var le1 := new OrdinalEncoder('region_name');
   df := le1.FitTransform(df);
   
-  var le2 := new LabelEncoder('federal_district');
+  var le2 := new OrdinalEncoder('federal_district');
   df := le2.FitTransform(df);
   
   var features := ['lat', 'lon', 'region_name', 'federal_district'];

Release/pabcversion.txt

@@ -1 +1 @@
-3.11.1.3815
+3.11.1.3821

ReleaseGenerators/PascalABCNET_version.nsh

@@ -1 +1 @@
-!define VERSION '3.11.1.3815'
+!define VERSION '3.11.1.3821'

TestSuite/_MachineLearning/Core/030_labelencoder_first_appearance.pas

@@ -6,7 +6,7 @@ begin
   df.AddStrColumn('Region', Arr('South', 'North', 'South', 'East'));
   df := df.SetCategorical(['Region']);
 
-  var enc := new LabelEncoder('Region');
+  var enc := new OrdinalEncoder('Region');
   enc.Fit(df);
 
   var res := enc.Transform(df);

TestSuite/_MachineLearning/Core/041_labelencoder_transform_type_drift_raises_contract_error.pas

@@ -10,7 +10,7 @@ begin
   testDf.AddIntColumn('City', Arr(1, 2, 3));
   testDf := testDf.SetCategorical(['City']);
 
-  var enc := new LabelEncoder('City');
+  var enc := new OrdinalEncoder('City');
   enc.Fit(trainDf);
 
   CheckRaises(procedure -> begin

bin/Lib/MLABC.pas

@@ -120,7 +120,7 @@ type
   Inspection = InspectionML.Inspection;
   
   IPreprocessor = PreprocessorABC.IPreprocessor;
-  LabelEncoder = PreprocessorABC.LabelEncoder;
+  OrdinalEncoder = PreprocessorABC.OrdinalEncoder;
   OneHotEncoder = PreprocessorABC.OneHotEncoder;
   ImputeStrategy = PreprocessorABC.ImputeStrategy;
   Imputer = PreprocessorABC.Imputer;

bin/Lib/MLDatasets.pas

@@ -1157,6 +1157,7 @@ begin
     ArgumentNullError(ER_ARG_NULL, 'name');
 
   var baseDir := PascalABCDirectory + 'Files\Datasets\';
+  //var baseDir := 'C:\Program Files (x86)\PascalABC.NET\Files\Datasets\';
 
   var metaPath := baseDir + name + '.meta';
   var csvPath  := baseDir + name + '.csv';

bin/Lib/MLModelsABC.pas

@@ -25,6 +25,14 @@ unit MLModelsABC;
 // См. статистическую политику в модуле MLABC.
 // =============================================================
 
+{
+  Производительность
+  
+  DecisionTreeRegressor.Fit - 340 мс против 156 мс в Питоне при той же точности
+  GradientBoostingRegressor.Fit - 5500 мс против 5500 мс в Питоне
+  RandomForestRegressor.Fit - 1280 мс против 480 мс в Питоне 
+}
+
 interface
 
 uses MLCoreABC;
@@ -432,6 +440,8 @@ type
     Found: boolean;
     Feature: integer;
     Threshold: real;
+    LeftCount: integer;
+    LeftOrderSize: integer;
     WeightedScore: real;
     
     static function Invalid: RegSplitResult;
@@ -439,6 +449,8 @@ type
       Result.Found := false;
       Result.Feature := -1;
       Result.Threshold := 0.0;
+      Result.LeftCount := 0;
+      Result.LeftOrderSize := 0;
       Result.WeightedScore := real.PositiveInfinity;
     end;
   end;
@@ -708,7 +720,8 @@ type
   private
     fLeafL2: real;
     fSortedOrders: array of array of integer;
-    fSortedValues: array of array of real;
+    fUseSortedOrdersAsRoot: boolean;
+    fRowWeights: array of integer;
       
     fVisitMarks: array of integer;
     fVisitId: integer;
@@ -726,12 +739,14 @@ type
     function FindBestSplitReg(X: Matrix; y: Vector; nodeOrders: array of array of integer): RegSplitResult;
     procedure BuildSortedOrders(X: Matrix; indices: array of integer);
     function BuildInitialNodeOrders(indices: array of integer): array of array of integer;
-    procedure SplitNodeOrders(X: Matrix; nodeOrders: array of array of integer; feature: integer; threshold: real;
+    procedure SplitNodeOrders(nodeOrders: array of array of integer; feature: integer; leftCount, leftOrderSize: integer;
       var leftOrders, rightOrders: array of array of integer);
     function BuildMembershipMask(rowCount: integer; indices: array of integer): array of boolean;
     procedure ComputeNodeStats(yData: array of real; indices: array of integer; var sumAll, sumSqAll: real);
     function WeightedVariance(n, leftCount: integer; leftSum, leftSumSq, sumAll, sumSqAll: real): real;
     function GetFeatureSubset(p: integer): array of integer;
+    function SampleWeight(rowIndex: integer): integer;
+    function TotalWeight(indices: array of integer): integer;
     
 /// Проверяет, является ли узел "чистым".
 /// Для регрессии это означает, что все значения y одинаковы
@@ -759,6 +774,16 @@ type
 /// Возвращает вектор вещественных значений.
     function Predict(X: Matrix): Vector; override;
     
+/// Внутренний hook для ансамблей: позволяет переиспользовать
+/// уже отсортированные порядки строк по признакам.
+/// Обычному пользовательскому коду не нужен.
+    procedure SetPreSortedOrders(sortedOrders: array of array of integer);
+    
+/// Внутренний hook для bootstrap-подвыборок с повторами.
+/// Передаются уже готовые сортированные порядки именно для корневого узла.
+    procedure SetPreSortedRootOrders(sortedOrders: array of array of integer);
+    procedure SetBootstrapRootOrders(sortedOrders: array of array of integer; rowWeights: array of integer);
+
 /// Копирует только конфигурацию модели (без обученного состояния).
 /// Используется для создания независимых экземпляров модели.
     function Clone: IModel; override;
@@ -4022,7 +4047,7 @@ function DecisionTreeRegressor.BuildTreeNode(X: Matrix; y: Vector;
   nodeOrders: array of array of integer; depth: integer): DecisionTreeNode;
 begin
   var indices := nodeOrders[0];
-  var n := indices.Length;
+  var n := TotalWeight(indices);
   
   if (fMaxDepth >= 0) and (depth >= fMaxDepth) then
     exit(LeafNode(LeafValue(y, indices)));
@@ -4048,12 +4073,13 @@ begin
     exit(LeafNode(LeafValue(y, indices)));
 
   var leftOrders, rightOrders: array of array of integer;
-  SplitNodeOrders(X, nodeOrders, split.Feature, split.Threshold, leftOrders, rightOrders);
+  SplitNodeOrders(nodeOrders, split.Feature, split.LeftCount, split.LeftOrderSize, leftOrders, rightOrders);
   var leftArr := leftOrders[0];
   var rightArr := rightOrders[0];
+  var rightCount := n - split.LeftCount;
 
-  if (leftArr.Length < fMinSamplesLeaf) or
-     (rightArr.Length < fMinSamplesLeaf) then
+  if (split.LeftCount < fMinSamplesLeaf) or
+     (rightCount < fMinSamplesLeaf) then
     exit(LeafNode(LeafValue(y, indices)));
 
   var delta := parentVar - split.WeightedScore;
@@ -4095,7 +4121,7 @@ begin
   Result := RegSplitResult.Invalid;
 
   var indices := nodeOrders[0];
-  var n := indices.Length;
+  var n := TotalWeight(indices);
   if n < 2 then
     exit;
 
@@ -4108,6 +4134,11 @@ begin
   var bestScore := real.PositiveInfinity;
   var bestFeature := -1;
   var bestThreshold := 0.0;
+  var bestLeftCount := 0;
+  var bestLeftOrderSize := 0;
+  var invN := 1.0 / n;
+  var minLeaf := fMinSamplesLeaf;
+  var maxLeftCount := n - minLeaf;
 
   var features := GetFeatureSubset(X.ColCount);
   
@@ -4117,63 +4148,61 @@ begin
     var order := nodeOrders[j];
     var orderLen := order.Length;
     
-    var leftCount := 0;
-    var leftSum := 0.0;
-    var leftSumSq := 0.0;
+    if orderLen < 2 then
+      continue;
 
-    var prevValue := 0.0;
-    var firstIncluded := true;
+    var firstIdx := order[0];
+    var firstWeight := SampleWeight(firstIdx);
+    var leftCount := firstWeight;
+    var leftSum := firstWeight * yData[firstIdx];
+    var leftSumSq := firstWeight * yData[firstIdx] * yData[firstIdx];
+    var prevValue := xData[firstIdx, j];
 
-    for var i := 0 to orderLen - 1 do
+    for var i := 1 to orderLen - 1 do
     begin
+      if leftCount > maxLeftCount then
+        break;
+        
       var idx := order[i];
       var xCur := xData[idx, j];
-      var yCur := yData[idx];
       
-      if not firstIncluded then
+      if (leftCount >= minLeaf) and (prevValue <> xCur) then
       begin
         var rightCount := n - leftCount;
-
-        if (leftCount >= fMinSamplesLeaf) and
-           (rightCount >= fMinSamplesLeaf) and
-           (prevValue <> xCur) then
-        begin
-          var rightCountReal := rightCount;
-          var leftMean := leftSum / leftCount;
-          var leftVar := leftSumSq / leftCount - leftMean * leftMean;
-
         var rightSum := sumAll - leftSum;
         var rightSumSq := sumSqAll - leftSumSq;
-          var rightMean := rightSum / rightCount;
-          var rightVar := rightSumSq / rightCount - rightMean * rightMean;
         
-          if leftVar < 0 then
-            leftVar := 0.0;
-          if rightVar < 0 then
-            rightVar := 0.0;
+        var leftScore := leftSumSq - (leftSum * leftSum) / leftCount;
+        var rightScore := rightSumSq - (rightSum * rightSum) / rightCount;
+        var weighted := (leftScore + rightScore) * invN;
         
-          var weighted := (leftCount * leftVar + rightCountReal * rightVar) / n;
+        if weighted < 0 then
+          weighted := 0.0;
 
         if weighted < bestScore then
         begin
           bestScore := weighted;
           bestFeature := j;
           bestThreshold := (prevValue + xCur) * 0.5;
-          end;
+          bestLeftCount := leftCount;
+          bestLeftOrderSize := i;
         end;
       end;
 
-      leftCount += 1;
-      leftSum += yCur;
-      leftSumSq += yCur * yCur;
+      var wCur := SampleWeight(idx);
+      var yCur := yData[idx];
+      leftCount += wCur;
+      leftSum += wCur * yCur;
+      leftSumSq += wCur * yCur * yCur;
       prevValue := xCur;
-      firstIncluded := false;
     end;
   end;
 
   Result.Found := bestFeature <> -1;
   Result.Feature := bestFeature;
   Result.Threshold := bestThreshold;
+  Result.LeftCount := bestLeftCount;
+  Result.LeftOrderSize := bestLeftOrderSize;
   Result.WeightedScore := bestScore;
 end;
 
@@ -4190,14 +4219,103 @@ type
     end;
   end;
 
+function BuildPreSortedOrders(X: Matrix): array of array of integer;
+begin
+  var p := X.ColCount;
+  var n := X.RowCount;
+  var xData := X.Data;
+  
+  SetLength(Result, p);
+  
+  for var j := 0 to p - 1 do
+  begin
+    var pairs: array of SortPair;
+    SetLength(pairs, n);
+    
+    for var i := 0 to n - 1 do
+    begin
+      pairs[i].Value := xData[i, j];
+      pairs[i].Index := i;
+    end;
+    
+    System.Array.Sort(pairs, new SortPairComparer);
+    
+    Result[j] := new integer[n];
+    for var i := 0 to n - 1 do
+      Result[j][i] := pairs[i].Index;
+  end;
+end;
+
+function BuildRowCounts(rows: array of integer; rowCount: integer): array of integer;
+begin
+  Result := new integer[rowCount];
+  
+  for var i := 0 to rows.Length - 1 do
+    Result[rows[i]] += 1;
+end;
+
+function BuildSortedOrdersFromCounts(
+  fullSortedOrders: array of array of integer;
+  rowCounts: array of integer
+): array of array of integer;
+begin
+  var p := Length(fullSortedOrders);
+  SetLength(Result, p);
+  
+  var total := 0;
+  for var i := 0 to rowCounts.Length - 1 do
+    total += rowCounts[i];
+  
+  for var j := 0 to p - 1 do
+  begin
+    Result[j] := new integer[total];
+    var k := 0;
+    
+    foreach var idx in fullSortedOrders[j] do
+      for var rep := 1 to rowCounts[idx] do
+      begin
+        Result[j][k] := idx;
+        k += 1;
+      end;
+  end;
+end;
+
+function BuildUniqueOrdersFromCounts(
+  fullSortedOrders: array of array of integer;
+  rowCounts: array of integer
+): array of array of integer;
+begin
+  var p := Length(fullSortedOrders);
+  SetLength(Result, p);
+  
+  var total := 0;
+  for var i := 0 to rowCounts.Length - 1 do
+    if rowCounts[i] > 0 then
+      total += 1;
+  
+  for var j := 0 to p - 1 do
+  begin
+    Result[j] := new integer[total];
+    var k := 0;
+    
+    foreach var idx in fullSortedOrders[j] do
+      if rowCounts[idx] > 0 then
+      begin
+        Result[j][k] := idx;
+        k += 1;
+      end;
+  end;
+end;
+
 procedure DecisionTreeRegressor.BuildSortedOrders(X: Matrix; indices: array of integer);
 begin
+  // Обычный путь для отдельного дерева:
+  // построить сортировку только по реально используемым строкам.
   var p := X.ColCount;
   var n := indices.Length;
   var xData := X.Data;
   
   SetLength(fSortedOrders, p);
-  SetLength(fSortedValues, p);
   
   for var j := 0 to p - 1 do
   begin
@@ -4214,14 +4332,10 @@ begin
     System.Array.Sort(pairs,new SortPairComparer);
     
     fSortedOrders[j] := new integer[n];
-    fSortedValues[j] := new real[n];
     
     for var i := 0 to n - 1 do
-    begin
-      fSortedValues[j][i] := pairs[i].Value;
       fSortedOrders[j][i] := pairs[i].Index;
   end;
-  end;
 end;
 
 function DecisionTreeRegressor.BuildInitialNodeOrders(indices: array of integer): array of array of integer;
@@ -4252,13 +4366,13 @@ begin
   end;
 end;
 
-procedure DecisionTreeRegressor.SplitNodeOrders(X: Matrix;
+procedure DecisionTreeRegressor.SplitNodeOrders(
   nodeOrders: array of array of integer;
   feature: integer;
-  threshold: real;
+  leftCount: integer;
+  leftOrderSize: integer;
   var leftOrders, rightOrders: array of array of integer);
 begin
-  var xData := X.Data;
   var p := Length(nodeOrders);
   
   fVisitId += 1;
@@ -4266,31 +4380,33 @@ begin
   
   var splitArr := nodeOrders[feature];
   var splitLen := splitArr.Length;
+  var rightCount := splitLen - leftOrderSize;
+  var markLeft := leftOrderSize <= rightCount;
   
-  var leftCount := 0;
-  
-  for var i := 0 to splitLen - 1 do
-  begin
-    var idx := splitArr[i];
-    
-    if xData[idx, feature] <= threshold then
-    begin
-      fVisitMarks[idx] := mark;
-      leftCount += 1;
-    end;
-  end;
-
-  var rightCount := splitLen - leftCount;
+  if markLeft then
+    for var i := 0 to leftOrderSize - 1 do
+      fVisitMarks[splitArr[i]] := mark
+  else
+    for var i := leftOrderSize to splitLen - 1 do
+      fVisitMarks[splitArr[i]] := mark;
 
   SetLength(leftOrders, p);
   SetLength(rightOrders, p);
 
+  leftOrders[feature] := new integer[leftOrderSize];
+  rightOrders[feature] := new integer[rightCount];
+  System.Array.Copy(splitArr, 0, leftOrders[feature], 0, leftOrderSize);
+  System.Array.Copy(splitArr, leftOrderSize, rightOrders[feature], 0, rightCount);
+
   for var j := 0 to p - 1 do
   begin
+    if j = feature then
+      continue;
+    
     var src := nodeOrders[j];
     var n := src.Length;
     
-    var left := new integer[leftCount];
+    var left := new integer[leftOrderSize];
     var right := new integer[rightCount];
     
     var li := 0;
@@ -4300,6 +4416,8 @@ begin
     begin
       var idx := src[k];
       
+      if markLeft then
+      begin
         if fVisitMarks[idx] = mark then
         begin
           left[li] := idx;
@@ -4310,6 +4428,20 @@ begin
           right[ri] := idx;
           ri += 1;
         end;
+      end
+      else
+      begin
+        if fVisitMarks[idx] = mark then
+        begin
+          right[ri] := idx;
+          ri += 1;
+        end
+        else
+        begin
+          left[li] := idx;
+          li += 1;
+        end;
+      end;
     end;
 
     leftOrders[j] := left;
@@ -4331,9 +4463,11 @@ begin
 
   for var i := 0 to indices.Length - 1 do
   begin
-    var v := yData[indices[i]];
-    sumAll += v;
-    sumSqAll += v * v;
+    var idx := indices[i];
+    var w := SampleWeight(idx);
+    var v := yData[idx];
+    sumAll += w * v;
+    sumSqAll += w * v * v;
   end;
 end;
 
@@ -4380,6 +4514,48 @@ begin
   end;
 end;
 
+procedure DecisionTreeRegressor.SetPreSortedOrders(sortedOrders: array of array of integer);
+begin
+  // Внутренний fast-path для ансамблей:
+  // используем готовую полную сортировку X и затем фильтруем её по fRowIndices.
+  fSortedOrders := sortedOrders;
+  fRowWeights := nil;
+  fUseSortedOrdersAsRoot := false;
+end;
+
+procedure DecisionTreeRegressor.SetPreSortedRootOrders(sortedOrders: array of array of integer);
+begin
+  // Внутренний fast-path для bootstrap-выборок с повторами:
+  // сортировка уже соответствует корневому узлу текущего дерева.
+  fSortedOrders := sortedOrders;
+  fRowWeights := nil;
+  fUseSortedOrdersAsRoot := true;
+end;
+
+procedure DecisionTreeRegressor.SetBootstrapRootOrders(
+  sortedOrders: array of array of integer;
+  rowWeights: array of integer);
+begin
+  fSortedOrders := sortedOrders;
+  fRowWeights := rowWeights;
+  fUseSortedOrdersAsRoot := true;
+end;
+
+function DecisionTreeRegressor.SampleWeight(rowIndex: integer): integer;
+begin
+  if fRowWeights = nil then
+    Result := 1
+  else
+    Result := fRowWeights[rowIndex];
+end;
+
+function DecisionTreeRegressor.TotalWeight(indices: array of integer): integer;
+begin
+  Result := 0;
+  for var i := 0 to indices.Length - 1 do
+    Result += SampleWeight(indices[i]);
+end;
+
 //==============================
 //    DecisionTreeClassifier
 //==============================
@@ -4571,7 +4747,7 @@ end;
 
 function DecisionTreeRegressor.LeafValue(y: Vector; indices: array of integer): real;
 begin
-  var n := indices.Length;
+  var n := TotalWeight(indices);
 
   if n = 0 then
     exit(0.0);  // безопасный fallback, не должен происходить
@@ -4582,7 +4758,7 @@ begin
   var sum := 0.0;
 
   foreach var idx in indices do
-    sum += y[idx];
+    sum += SampleWeight(idx) * y[idx];
 
   var denom: real;
 
@@ -4637,11 +4813,15 @@ begin
   if indices = nil then
     indices := Arr(0..X.RowCount - 1);
 
+  if (fSortedOrders = nil) or (Length(fSortedOrders) = 0) then
     BuildSortedOrders(X, indices);
   
   SetLength(fVisitMarks, X.RowCount);
   fVisitId := 0;
 
+  if fUseSortedOrdersAsRoot then
+    fRoot := BuildTreeNode(X, y, fSortedOrders, 0)
+  else
     fRoot := BuildTreeNew(X, y, indices, 0);
   
   var s := fFeatureImportances.Sum;
@@ -4653,7 +4833,8 @@ begin
   
   fRowIndices := nil;
   fSortedOrders := nil;
-  fSortedValues := nil;
+  fUseSortedOrdersAsRoot := false;
+  fRowWeights := nil;
   
   fVisitMarks := nil;
 
@@ -4830,6 +5011,7 @@ begin
   fFeatureCount := p;
 
   SetLength(fTrees, fNTrees);
+  var fullSortedOrders := BuildPreSortedOrders(X);
 
   // --- OOB buffers (regression) ---
   var oobSum: Vector := nil;
@@ -4859,7 +5041,10 @@ begin
     var rows: array of integer;
     BootstrapRowIndices(n, rows);
     
-    tree.SetRowIndices(rows);
+    var rowCounts := BuildRowCounts(rows, n);
+    var bootSortedOrders := BuildUniqueOrdersFromCounts(fullSortedOrders, rowCounts);
+    tree.SetBootstrapRootOrders(bootSortedOrders, rowCounts);
+    
     tree.Fit(X, y);
 
     // --- OOB accumulate ---
@@ -5730,6 +5915,9 @@ begin
   for var i := 0 to nTrain - 1 do
     yPredTrain[i] := fInitValue;
   
+  var residuals := new Vector(nTrain);
+  var preSortedOrders := BuildPreSortedOrders(XTrain);
+
   // --- OOB logic ---
   var useSubsample := fSubsample < 1.0;
   var useOOB :=
@@ -5754,8 +5942,7 @@ begin
   for var m := 0 to fNEstimators - 1 do
   begin
     // 1. residuals
-    var r := new Vector(nTrain);
-    ComputePseudoResiduals(yTrain, yPredTrain, r);
+    ComputePseudoResiduals(yTrain, yPredTrain, residuals);
 
     var stageSeed := fRng.Next(integer.MaxValue);
 
@@ -5766,6 +5953,7 @@ begin
       fLeafL2,
       stageSeed
     );
+    tree.SetPreSortedOrders(preSortedOrders);
 
     // --- subsample ---
     var rows: array of integer := nil;
@@ -5784,7 +5972,7 @@ begin
       end;
     end;
     
-    tree.Fit(XTrain, r);
+    tree.Fit(XTrain, residuals);
     fEstimators.Add(tree);
 
     var deltaTrain := tree.Predict(XTrain);

bin/Lib/PreprocessorABC.pas

@@ -48,14 +48,14 @@ type
 /// Пропущенные значения (NA) игнорируются при обучении
 /// и сохраняются как пропуски при преобразовании.
 /// Работает только со строковыми столбцами и предназначен для признаков.
-/// Не должен применяться к целевому столбцу (target).
-  LabelEncoder = class(IPreprocessor, IColumnBoundStep)
+/// Не должен применяться к целевому столбцу (target)
+  OrdinalEncoder = class(IPreprocessor, IColumnBoundStep)
   private
     col: string;
     mapping: Dictionary<string, integer>;
     fitted: boolean;
   public
-    /// Создаёт LabelEncoder для указанного столбца
+    /// Создаёт OrdinalEncoder для указанного столбца
     constructor Create(column: string);
   
 /// Определяет множество категорий столбца и сохраняет их числовое кодирование.
@@ -66,8 +66,10 @@ type
 ///   • отображение категорий НЕ копируется методом Clone    
     function Fit(df: DataFrame): IPreprocessor;
     
-    /// Заменяет категории их числовыми кодами
-    /// Возвращает новый DataFrame
+    /// Заменяет категории их числовыми кодами.
+    /// Неизвестные категории кодируются значением -1.
+    /// Пропущенные значения сохраняются как пропуски.
+    /// Возвращает новый DataFrame.
     function Transform(df: DataFrame): DataFrame;
     
     /// Выполняет Fit и Transform последовательно
@@ -174,14 +176,14 @@ implementation
 uses MLExceptions;
 
 const
-  ER_LABELENCODER_NO_COLUMN =
-    'LabelEncoder: столбец не указан!!LabelEncoder: column not specified';
-  ER_LABELENCODER_NOT_STRING =
-    'LabelEncoder: столбец "{0}" не является строковым!!' +
-    'LabelEncoder: column "{0}" is not string';
-  ER_LABELENCODER_UNSEEN_CATEGORY =
-    'LabelEncoder: неизвестная категория "{0}"!!' +
-    'LabelEncoder: unseen category "{0}"';
+  ER_ORDINALENCODER_NO_COLUMN =
+    'OrdinalEncoder: столбец не указан!!OrdinalEncoder: column not specified';
+  ER_ORDINALENCODER_NOT_STRING =
+    'OrdinalEncoder: столбец "{0}" не является строковым!!' +
+    'OrdinalEncoder: column "{0}" is not string';
+  ER_ORDINALENCODER_UNSEEN_CATEGORY =
+    'OrdinalEncoder: неизвестная категория "{0}"!!' +
+    'OrdinalEncoder: unseen category "{0}"';
   ER_ONEHOT_NO_COLUMN =
     'OneHotEncoder: столбец не указан!!OneHotEncoder: column not specified';
   ER_ONEHOT_NOT_STRING =
@@ -229,19 +231,19 @@ const
     
   
 //-----------------------------
-//        LabelEncoder
+//        OrdinalEncoder
 //-----------------------------
 
-constructor LabelEncoder.Create(column: string);
+constructor OrdinalEncoder.Create(column: string);
 begin
   if column = '' then
-    ArgumentError(ER_LABELENCODER_NO_COLUMN);
+    ArgumentError(ER_ORDINALENCODER_NO_COLUMN);
 
   col := column;
   fitted := false;
 end;
 
-function LabelEncoder.Fit(df: DataFrame): IPreprocessor;
+function OrdinalEncoder.Fit(df: DataFrame): IPreprocessor;
 begin
   if df = nil then
     ArgumentNullError(ER_ARG_NULL, 'df');
@@ -252,7 +254,7 @@ begin
   var idx := df.Schema.IndexOf(col);
 
   if df.Schema.ColumnTypeAt(idx) <> ColumnType.ctStr then
-    Error(ER_LABELENCODER_NOT_STRING, col);
+    Error(ER_ORDINALENCODER_NOT_STRING, col);
 
   mapping := new Dictionary<string, integer>;
 
@@ -275,7 +277,7 @@ begin
   Result := Self;
 end;
 
-function LabelEncoder.Transform(df: DataFrame): DataFrame;
+function OrdinalEncoder.Transform(df: DataFrame): DataFrame;
 begin
   if not fitted then
     NotFittedError(ER_FIT_NOT_CALLED);
@@ -299,10 +301,11 @@ begin
     begin
       var s := cur.Str(idx);
 
-      if not mapping.ContainsKey(s) then
-        Error(ER_LABELENCODER_UNSEEN_CATEGORY, s);
-
+      if s not in mapping then
+        data[row] := -1
+      else
         data[row] := mapping[s];
+      
       valid[row] := True;
     end;
 
@@ -328,20 +331,20 @@ begin
   Result := res.SetCategorical(catCols.ToArray);
 end;
 
-function LabelEncoder.FitTransform(df: DataFrame): DataFrame;
+function OrdinalEncoder.FitTransform(df: DataFrame): DataFrame;
 begin
   Fit(df);
   Result := Transform(df);
 end;
 
-function LabelEncoder.ToString: string;
+function OrdinalEncoder.ToString: string;
 begin
-  Result := 'LabelEncoder(' + col + ')';
+  Result := 'OrdinalEncoder(' + col + ')';
 end;
 
-function LabelEncoder.Clone: IPreprocessor;
+function OrdinalEncoder.Clone: IPreprocessor;
 begin
-  Result := new LabelEncoder(col);
+  Result := new OrdinalEncoder(col);
 end;
 
 procedure AppendAllColumnsExcept(