Changeset 14345 for trunk/sources/HeuristicLab.Algorithms.DataAnalysis/3.4/RandomForest

Timestamp:

10/21/16 17:59:36 (8 years ago)

Author:

gkronber

Message:

#2690: implemented methods to generate symbolic expression tree solutions for decision tree models (random forest and gradient boosted) as well as views which make it possible to inspect each of the individual trees in a GBT and RF solution

Location:

trunk/sources/HeuristicLab.Algorithms.DataAnalysis/3.4/RandomForest

Files:

• RandomForestClassificationSolution.cs (modified) (1 diff)
• RandomForestModel.cs (modified) (4 diffs)
• RandomForestRegressionSolution.cs (modified) (2 diffs)

trunk/sources/HeuristicLab.Algorithms.DataAnalysis/3.4/RandomForest/RandomForestClassificationSolution.cs

@@ -38 +38 @@
     }
 
+    public int NumberOfTrees {
+      get { return Model.NumberOfTrees; }
+    }
+
     [StorableConstructor]
     private RandomForestClassificationSolution(bool deserializing) : base(deserializing) { }

trunk/sources/HeuristicLab.Algorithms.DataAnalysis/3.4/RandomForest/RandomForestModel.cs

@@ -25 +25 @@
 using HeuristicLab.Common;
 using HeuristicLab.Core;
+using HeuristicLab.Encodings.SymbolicExpressionTreeEncoding;
 using HeuristicLab.Persistence.Default.CompositeSerializers.Storable;
 using HeuristicLab.Problems.DataAnalysis;
+using HeuristicLab.Problems.DataAnalysis.Symbolic;
 
 namespace HeuristicLab.Algorithms.DataAnalysis {
@@ -49 +51 @@
     }
 
+    public int NumberOfTrees {
+      get { return nTrees; }
+    }
 
     // instead of storing the data of the model itself
@@ -64 +69 @@
     [Storable]
     private double m;
-
 
     [StorableConstructor]
@@ -197 +201 @@
     }
 
+    public ISymbolicExpressionTree ExtractTree(int treeIdx) {
+      // hoping that the internal representation of alglib is stable
+
+      // TREE FORMAT
+      // W[Offs]      -   size of sub-array (for the tree)
+      //     node info:
+      // W[K+0]       -   variable number        (-1 for leaf mode)
+      // W[K+1]       -   threshold              (class/value for leaf node)
+      // W[K+2]       -   ">=" branch index      (absent for leaf node)
+
+      // skip irrelevant trees
+      int offset = 0;
+      for (int i = 0; i < treeIdx - 1; i++) {
+        offset = offset + (int)Math.Round(randomForest.innerobj.trees[offset]);
+      }
+
+      var constSy = new Constant();
+      var varCondSy = new VariableCondition() { IgnoreSlope = true };
+
+      var node = CreateRegressionTreeRec(randomForest.innerobj.trees, offset, offset + 1, constSy, varCondSy);
+
+      var startNode = new StartSymbol().CreateTreeNode();
+      startNode.AddSubtree(node);
+      var root = new ProgramRootSymbol().CreateTreeNode();
+      root.AddSubtree(startNode);
+      return new SymbolicExpressionTree(root);
+    }
+
+    private ISymbolicExpressionTreeNode CreateRegressionTreeRec(double[] trees, int offset, int k, Constant constSy, VariableCondition varCondSy) {
+
+      // alglib source for evaluation of one tree (dfprocessinternal)
+      // offs = 0
+      //
+      // Set pointer to the root
+      //
+      // k = offs + 1;
+      // 
+      // //
+      // // Navigate through the tree
+      // //
+      // while (true) {
+      //   if ((double)(df.trees[k]) == (double)(-1)) {
+      //     if (df.nclasses == 1) {
+      //       y[0] = y[0] + df.trees[k + 1];
+      //     } else {
+      //       idx = (int)Math.Round(df.trees[k + 1]);
+      //       y[idx] = y[idx] + 1;
+      //     }
+      //     break;
+      //   }
+      //   if ((double)(x[(int)Math.Round(df.trees[k])]) < (double)(df.trees[k + 1])) {
+      //     k = k + innernodewidth;
+      //   } else {
+      //     k = offs + (int)Math.Round(df.trees[k + 2]);
+      //   }
+      // }
+
+      if ((double)(trees[k]) == (double)(-1)) {
+        var constNode = (ConstantTreeNode)constSy.CreateTreeNode();
+        constNode.Value = trees[k + 1];
+        return constNode;
+      } else {
+        var condNode = (VariableConditionTreeNode)varCondSy.CreateTreeNode();
+        condNode.VariableName = AllowedInputVariables[(int)Math.Round(trees[k])];
+        condNode.Threshold = trees[k + 1];
+        condNode.Slope = double.PositiveInfinity;
+
+        var left = CreateRegressionTreeRec(trees, offset, k + 3, constSy, varCondSy);
+        var right = CreateRegressionTreeRec(trees, offset, offset + (int)Math.Round(trees[k + 2]), constSy, varCondSy);
+
+        condNode.AddSubtree(left); // not 100% correct because interpreter uses: if(x <= thres) left() else right() and RF uses if(x < thres) left() else right() (see above)
+        condNode.AddSubtree(right);
+        return condNode;
+      }
+    }
+
 
     public IRegressionSolution CreateRegressionSolution(IRegressionProblemData problemData) {

trunk/sources/HeuristicLab.Algorithms.DataAnalysis/3.4/RandomForest/RandomForestRegressionSolution.cs

@@ -20 +20 @@
 #endregion
 
+using System;
 using HeuristicLab.Common;
 using HeuristicLab.Core;
 using HeuristicLab.Persistence.Default.CompositeSerializers.Storable;
 using HeuristicLab.Problems.DataAnalysis;
+using HeuristicLab.Problems.DataAnalysis.Symbolic.Regression;
 
 namespace HeuristicLab.Algorithms.DataAnalysis {
@@ -36 +38 @@
       get { return (IRandomForestModel)base.Model; }
       set { base.Model = value; }
+    }
+
+    public int NumberOfTrees {
+      get { return Model.NumberOfTrees; }
     }