DecisionTreeNode.h

Go to the documentation of this file.

// @(#)root/tmva $Id$
// Author: Andreas Hoecker, Joerg Stelzer, Helge Voss, Kai Voss, Eckhard von Toerne

/**********************************************************************************
 * Project: TMVA - a Root-integrated toolkit for multivariate data analysis       *
 * Package: TMVA                                                                  *
 * Class  : DecisionTreeNode                                                      *
 * Web    : http://tmva.sourceforge.net                                           *
 *                                                                                *
 * Description:                                                                   *
 *      Node for the Decision Tree                                                *
 *                                                                                *
 * Authors (alphabetical):                                                        *
 *      Andreas Hoecker <Andreas.Hocker@cern.ch> - CERN, Switzerland              *
 *      Helge Voss      <Helge.Voss@cern.ch>     - MPI-K Heidelberg, Germany      *
 *      Kai Voss        <Kai.Voss@cern.ch>       - U. of Victoria, Canada         *
 *      Eckhard von Toerne <evt@physik.uni-bonn.de>  - U. of Bonn, Germany        *
 *                                                                                *
 * Copyright (c) 2009:                                                            *
 *      CERN, Switzerland                                                         *
 *      U. of Victoria, Canada                                                    *
 *      MPI-K Heidelberg, Germany                                                 *
*       U. of Bonn, Germany                                                       *
 *                                                                                *
 * Redistribution and use in source and binary forms, with or without             *
 * modification, are permitted according to the terms listed in LICENSE           *
 * (http://tmva.sourceforge.net/LICENSE)                                          *
 **********************************************************************************/

#ifndef ROOT_TMVA_DecisionTreeNode
#define ROOT_TMVA_DecisionTreeNode

//////////////////////////////////////////////////////////////////////////
//                                                                      //
// DecisionTreeNode                                                     //
//                                                                      //
// Node for the Decision Tree                                           //
//                                                                      //
//////////////////////////////////////////////////////////////////////////

#ifndef ROOT_TMVA_Node
#include "TMVA/Node.h"
#endif

#ifndef ROOT_TMVA_Version
#include "TMVA/Version.h"
#endif

#include <iostream>
#include <vector>
#include <map>
namespace TMVA {

   class DTNodeTrainingInfo
   {
   public:
      DTNodeTrainingInfo():fSampleMin(), 
                           fSampleMax(), 
                           fNodeR(0),fSubTreeR(0),fAlpha(0),fG(0),fNTerminal(0),
                           fNB(0),fNS(0),fSumTarget(0),fSumTarget2(0),fCC(0), 
                           fNSigEvents ( 0 ), fNBkgEvents ( 0 ),
                           fNEvents ( -1 ),
                           fNSigEvents_unweighted ( 0 ),
                           fNBkgEvents_unweighted ( 0 ),
                           fNEvents_unweighted ( 0 ),
                           fNSigEvents_unboosted ( 0 ),
                           fNBkgEvents_unboosted ( 0 ),
                           fNEvents_unboosted ( 0 ),
                           fSeparationIndex (-1 ),
                           fSeparationGain ( -1 )
      {
      }
      std::vector< Float_t >  fSampleMin; // the minima for each ivar of the sample on the node during training
      std::vector< Float_t >  fSampleMax; // the maxima for each ivar of the sample on the node during training
      Double_t fNodeR;           // node resubstitution estimate, R(t)
      Double_t fSubTreeR;        // R(T) = Sum(R(t) : t in ~T)
      Double_t fAlpha;           // critical alpha for this node
      Double_t fG;               // minimum alpha in subtree rooted at this node
      Int_t    fNTerminal;       // number of terminal nodes in subtree rooted at this node
      Double_t fNB;              // sum of weights of background events from the pruning sample in this node
      Double_t fNS;              // ditto for the signal events
      Float_t  fSumTarget;       // sum of weight*target  used for the calculatio of the variance (regression)
      Float_t  fSumTarget2;      // sum of weight*target^2 used for the calculatio of the variance (regression)
      Double_t fCC;  // debug variable for cost complexity pruning ..

      Float_t  fNSigEvents;      // sum of weights of signal event in the node
      Float_t  fNBkgEvents;      // sum of weights of backgr event in the node
      Float_t  fNEvents;         // number of events in that entered the node (during training)
      Float_t  fNSigEvents_unweighted;      // sum of signal event in the node
      Float_t  fNBkgEvents_unweighted;      // sum of backgr event in the node
      Float_t  fNEvents_unweighted;         // number of events in that entered the node (during training)
      Float_t  fNSigEvents_unboosted;      // sum of signal event in the node
      Float_t  fNBkgEvents_unboosted;      // sum of backgr event in the node
      Float_t  fNEvents_unboosted;         // number of events in that entered the node (during training)
      Float_t  fSeparationIndex; // measure of "purity" (separation between S and B) AT this node
      Float_t  fSeparationGain;  // measure of "purity", separation, or information gained BY this nodes selection

      // copy constructor
      DTNodeTrainingInfo(const DTNodeTrainingInfo& n) :
         fSampleMin(),fSampleMax(), // Samplemin and max are reset in copy constructor
         fNodeR(n.fNodeR), fSubTreeR(n.fSubTreeR),
         fAlpha(n.fAlpha), fG(n.fG),
         fNTerminal(n.fNTerminal),
         fNB(n.fNB), fNS(n.fNS),
         fSumTarget(0),fSumTarget2(0), // SumTarget reset in copy constructor
         fCC(0),
         fNSigEvents ( n.fNSigEvents ), fNBkgEvents ( n.fNBkgEvents ),
         fNEvents ( n.fNEvents ),
         fNSigEvents_unweighted ( n.fNSigEvents_unweighted ),
         fNBkgEvents_unweighted ( n.fNBkgEvents_unweighted ),
         fNEvents_unweighted ( n.fNEvents_unweighted ),
         fSeparationIndex( n.fSeparationIndex ),
         fSeparationGain ( n.fSeparationGain )
      { }
   };

   class Event;
   class MsgLogger;

   class DecisionTreeNode: public Node {

   public:

      // constructor of an essentially "empty" node floating in space
      DecisionTreeNode ();
      // constructor of a daughter node as a daughter of 'p'
      DecisionTreeNode (Node* p, char pos); 
    
      // copy constructor 
      DecisionTreeNode (const DecisionTreeNode &n, DecisionTreeNode* parent = NULL); 
      
      // destructor
      virtual ~DecisionTreeNode();

      virtual Node* CreateNode() const { return new DecisionTreeNode(); }

      inline void SetNFisherCoeff(Int_t nvars){fFisherCoeff.resize(nvars);}
      inline UInt_t GetNFisherCoeff() const { return fFisherCoeff.size();}
      // set fisher coefficients
      void SetFisherCoeff(Int_t ivar, Double_t coeff);      
      // get fisher coefficients
      Double_t GetFisherCoeff(Int_t ivar) const {return fFisherCoeff.at(ivar);}

      // test event if it decends the tree at this node to the right
      virtual Bool_t GoesRight( const Event & ) const;

      // test event if it decends the tree at this node to the left
      virtual Bool_t GoesLeft ( const Event & ) const;

      // set index of variable used for discrimination at this node
      void SetSelector( Short_t i) { fSelector = i; }
      // return index of variable used for discrimination at this node
      Short_t GetSelector() const { return fSelector; }

      // set the cut value applied at this node
      void  SetCutValue ( Float_t c ) { fCutValue  = c; }
      // return the cut value applied at this node
      Float_t GetCutValue ( void ) const { return fCutValue;  }

      // set true: if event variable > cutValue ==> signal , false otherwise
      void SetCutType( Bool_t t   ) { fCutType = t; }
      // return kTRUE: Cuts select signal, kFALSE: Cuts select bkg
      Bool_t GetCutType( void ) const { return fCutType; }

      // set node type: 1 signal node, -1 bkg leave, 0 intermediate Node
      void  SetNodeType( Int_t t ) { fNodeType = t;}
      // return node type: 1 signal node, -1 bkg leave, 0 intermediate Node
      Int_t GetNodeType( void ) const { return fNodeType; }

      //return  S/(S+B) (purity) at this node (from  training)
      Float_t GetPurity( void ) const { return fPurity;}
      //calculate S/(S+B) (purity) at this node (from  training)
      void SetPurity( void );

      //set the response of the node (for regression)
      void SetResponse( Float_t r ) { fResponse = r;}

      //return the response of the node (for regression)
      Float_t GetResponse( void ) const { return fResponse;}

      //set the RMS of the response of the node (for regression)
      void SetRMS( Float_t r ) { fRMS = r;}

      //return the RMS of the response of the node (for regression)
      Float_t GetRMS( void ) const { return fRMS;}

      // set the sum of the signal weights in the node
      void SetNSigEvents( Float_t s ) { fTrainInfo->fNSigEvents = s; }

      // set the sum of the backgr weights in the node
      void SetNBkgEvents( Float_t b ) { fTrainInfo->fNBkgEvents = b; }

      // set the number of events that entered the node (during training)
      void SetNEvents( Float_t nev ){ fTrainInfo->fNEvents =nev ; }

      // set the sum of the unweighted signal events in the node
      void SetNSigEvents_unweighted( Float_t s ) { fTrainInfo->fNSigEvents_unweighted = s; }

      // set the sum of the unweighted backgr events in the node
      void SetNBkgEvents_unweighted( Float_t b ) { fTrainInfo->fNBkgEvents_unweighted = b; }

      // set the number of unweighted events that entered the node (during training)
      void SetNEvents_unweighted( Float_t nev ){ fTrainInfo->fNEvents_unweighted =nev ; }

      // set the sum of the unboosted signal events in the node
      void SetNSigEvents_unboosted( Float_t s ) { fTrainInfo->fNSigEvents_unboosted = s; }

      // set the sum of the unboosted backgr events in the node
      void SetNBkgEvents_unboosted( Float_t b ) { fTrainInfo->fNBkgEvents_unboosted = b; }

      // set the number of unboosted events that entered the node (during training)
      void SetNEvents_unboosted( Float_t nev ){ fTrainInfo->fNEvents_unboosted =nev ; }

      // increment the sum of the signal weights in the node
      void IncrementNSigEvents( Float_t s ) { fTrainInfo->fNSigEvents += s; }

      // increment the sum of the backgr weights in the node
      void IncrementNBkgEvents( Float_t b ) { fTrainInfo->fNBkgEvents += b; }

      // increment the number of events that entered the node (during training)
      void IncrementNEvents( Float_t nev ){ fTrainInfo->fNEvents +=nev ; }

      // increment the sum of the signal weights in the node
      void IncrementNSigEvents_unweighted( ) { fTrainInfo->fNSigEvents_unweighted += 1; }

      // increment the sum of the backgr weights in the node
      void IncrementNBkgEvents_unweighted( ) { fTrainInfo->fNBkgEvents_unweighted += 1; }

      // increment the number of events that entered the node (during training)
      void IncrementNEvents_unweighted( ){ fTrainInfo->fNEvents_unweighted +=1 ; }

      // return the sum of the signal weights in the node
      Float_t GetNSigEvents( void ) const  { return fTrainInfo->fNSigEvents; }

      // return the sum of the backgr weights in the node
      Float_t GetNBkgEvents( void ) const  { return fTrainInfo->fNBkgEvents; }

      // return  the number of events that entered the node (during training)
      Float_t GetNEvents( void ) const  { return fTrainInfo->fNEvents; }

      // return the sum of unweighted signal weights in the node
      Float_t GetNSigEvents_unweighted( void ) const  { return fTrainInfo->fNSigEvents_unweighted; }

      // return the sum of unweighted backgr weights in the node
      Float_t GetNBkgEvents_unweighted( void ) const  { return fTrainInfo->fNBkgEvents_unweighted; }

      // return  the number of unweighted events that entered the node (during training)
      Float_t GetNEvents_unweighted( void ) const  { return fTrainInfo->fNEvents_unweighted; }

      // return the sum of unboosted signal weights in the node
      Float_t GetNSigEvents_unboosted( void ) const  { return fTrainInfo->fNSigEvents_unboosted; }

      // return the sum of unboosted backgr weights in the node
      Float_t GetNBkgEvents_unboosted( void ) const  { return fTrainInfo->fNBkgEvents_unboosted; }

      // return  the number of unboosted events that entered the node (during training)
      Float_t GetNEvents_unboosted( void ) const  { return fTrainInfo->fNEvents_unboosted; }


      // set the choosen index, measure of "purity" (separation between S and B) AT this node
      void SetSeparationIndex( Float_t sep ){ fTrainInfo->fSeparationIndex =sep ; }
      // return the separation index AT this node
      Float_t GetSeparationIndex( void ) const  { return fTrainInfo->fSeparationIndex; }

      // set the separation, or information gained BY this nodes selection
      void SetSeparationGain( Float_t sep ){ fTrainInfo->fSeparationGain =sep ; }
      // return the gain in separation obtained by this nodes selection
      Float_t GetSeparationGain( void ) const  { return fTrainInfo->fSeparationGain; }

      // printout of the node
      virtual void Print( std::ostream& os ) const;

      // recursively print the node and its daughters (--> print the 'tree')
      virtual void PrintRec( std::ostream&  os ) const;

      virtual void AddAttributesToNode(void* node) const;
      virtual void AddContentToNode(std::stringstream& s) const;

      // recursively clear the nodes content (S/N etc, but not the cut criteria)
      void ClearNodeAndAllDaughters();

      // get pointers to children, mother in the tree

      // return pointer to the left/right daughter or parent node
      inline virtual DecisionTreeNode* GetLeft( )   const { return dynamic_cast<DecisionTreeNode*>(fLeft); }
      inline virtual DecisionTreeNode* GetRight( )  const { return dynamic_cast<DecisionTreeNode*>(fRight); }
      inline virtual DecisionTreeNode* GetParent( ) const { return dynamic_cast<DecisionTreeNode*>(fParent); }

      // set pointer to the left/right daughter and parent node
      inline virtual void SetLeft  (Node* l) { fLeft   = dynamic_cast<DecisionTreeNode*>(l);} 
      inline virtual void SetRight (Node* r) { fRight  = dynamic_cast<DecisionTreeNode*>(r);} 
      inline virtual void SetParent(Node* p) { fParent = dynamic_cast<DecisionTreeNode*>(p);} 




      // the node resubstitution estimate, R(t), for Cost Complexity pruning
      inline void SetNodeR( Double_t r ) { fTrainInfo->fNodeR = r;    }
      inline Double_t GetNodeR( ) const  { return fTrainInfo->fNodeR; }

      // the resubstitution estimate, R(T_t), of the tree rooted at this node
      inline void SetSubTreeR( Double_t r ) { fTrainInfo->fSubTreeR = r;    }
      inline Double_t GetSubTreeR( ) const  { return fTrainInfo->fSubTreeR; }

      //                             R(t) - R(T_t)
      // the critical point alpha =  -------------
      //                              |~T_t| - 1
      inline void SetAlpha( Double_t alpha ) { fTrainInfo->fAlpha = alpha; }
      inline Double_t GetAlpha( ) const      { return fTrainInfo->fAlpha;  }

      // the minimum alpha in the tree rooted at this node
      inline void SetAlphaMinSubtree( Double_t g ) { fTrainInfo->fG = g;    }
      inline Double_t GetAlphaMinSubtree( ) const  { return fTrainInfo->fG; }

      // number of terminal nodes in the subtree rooted here
      inline void SetNTerminal( Int_t n ) { fTrainInfo->fNTerminal = n;    }
      inline Int_t GetNTerminal( ) const  { return fTrainInfo->fNTerminal; }

      // number of background/signal events from the pruning validation sample
      inline void SetNBValidation( Double_t b ) { fTrainInfo->fNB = b; }
      inline void SetNSValidation( Double_t s ) { fTrainInfo->fNS = s; }
      inline Double_t GetNBValidation( ) const  { return fTrainInfo->fNB; }
      inline Double_t GetNSValidation( ) const  { return fTrainInfo->fNS; }


      inline void SetSumTarget(Float_t t)  {fTrainInfo->fSumTarget = t; }
      inline void SetSumTarget2(Float_t t2){fTrainInfo->fSumTarget2 = t2; }

      inline void AddToSumTarget(Float_t t)  {fTrainInfo->fSumTarget += t; }
      inline void AddToSumTarget2(Float_t t2){fTrainInfo->fSumTarget2 += t2; }

      inline Float_t GetSumTarget()  const {return fTrainInfo? fTrainInfo->fSumTarget : -9999;}
      inline Float_t GetSumTarget2() const {return fTrainInfo? fTrainInfo->fSumTarget2: -9999;}


      // reset the pruning validation data
      void ResetValidationData( );

      // flag indicates whether this node is terminal
      inline Bool_t IsTerminal() const            { return fIsTerminalNode; }
      inline void SetTerminal( Bool_t s = kTRUE ) { fIsTerminalNode = s;    }
      void PrintPrune( std::ostream& os ) const ;
      void PrintRecPrune( std::ostream& os ) const;

      void     SetCC(Double_t cc);
      Double_t GetCC() const {return (fTrainInfo? fTrainInfo->fCC : -1.);}

      Float_t GetSampleMin(UInt_t ivar) const;
      Float_t GetSampleMax(UInt_t ivar) const;
      void     SetSampleMin(UInt_t ivar, Float_t xmin);
      void     SetSampleMax(UInt_t ivar, Float_t xmax);

      static bool fgIsTraining; // static variable to flag training phase in which we need fTrainInfo
      static UInt_t fgTmva_Version_Code;  // set only when read from weightfile 

   protected:

      static MsgLogger& Log();

      std::vector<Double_t>       fFisherCoeff;    // the fisher coeff (offset at the last element)

      Float_t  fCutValue;        // cut value appplied on this node to discriminate bkg against sig
      Bool_t   fCutType;         // true: if event variable > cutValue ==> signal , false otherwise
      Short_t  fSelector;        // index of variable used in node selection (decision tree)

      Float_t  fResponse;        // response value in case of regression
      Float_t  fRMS;             // response RMS of the regression node
      Int_t    fNodeType;        // Type of node: -1 == Bkg-leaf, 1 == Signal-leaf, 0 = internal
      Float_t  fPurity;          // the node purity

      Bool_t   fIsTerminalNode;    //! flag to set node as terminal (i.e., without deleting its descendants)

      mutable DTNodeTrainingInfo* fTrainInfo;

   private:

      virtual void ReadAttributes(void* node, UInt_t tmva_Version_Code = TMVA_VERSION_CODE );
      virtual Bool_t ReadDataRecord( std::istream& is, UInt_t tmva_Version_Code = TMVA_VERSION_CODE );
      virtual void ReadContent(std::stringstream& s);

      ClassDef(DecisionTreeNode,0) // Node for the Decision Tree 
   };
} // namespace TMVA

#endif