ExactBNdistance computes exactly the KL divergence betweens 2 BNs. More...

#include <exactBNdistance.h>

Inheritance diagram for gum::ExactBNdistance< GUM_SCALAR >:

Collaboration diagram for gum::ExactBNdistance< GUM_SCALAR >:

Public Member Functions
	ExactBNdistance (const IBayesNet< GUM_SCALAR > &P, const IBayesNet< GUM_SCALAR > &Q)
	constructor must give 2 BNs
	ExactBNdistance (const BNdistance< GUM_SCALAR > &kl)
	copy constructor
virtual	~ExactBNdistance ()
	destructor
Complexity	difficulty () const
	return KL::Complexity::Heavy,KL::Complexity::Difficult,KL::Complexity::Correct depending on the BNs p and q
Accessors to results. The first call do the computations. The
others do not.
double	klPQ ()
Size	errorPQ ()
double	klQP ()
Size	errorQP ()
double	hellinger ()
double	bhattacharya ()
double	jsd ()
const IBayesNet< GUM_SCALAR > &	p () const
const IBayesNet< GUM_SCALAR > &	q () const

Protected Member Functions
void	computeKL_ () final
void	process_ ()

Protected Attributes
const IBayesNet< GUM_SCALAR > &	p_
const IBayesNet< GUM_SCALAR > &	q_
GUM_SCALAR	klPQ_
GUM_SCALAR	klQP_
Size	errorPQ_
Size	errorQP_
GUM_SCALAR	hellinger_
GUM_SCALAR	bhattacharya_
GUM_SCALAR	jsd_

Private Member Functions
bool	_checkCompatibility_ () const

Private Attributes
Complexity	_difficulty_
bool	_done_

Detailed Description

template<typename GUM_SCALAR>
class gum::ExactBNdistance< GUM_SCALAR >

ExactBNdistance computes exactly the KL divergence betweens 2 BNs.

ExactBNdistance should be used only if difficulty() gives an estimation ( KL_CORRECT ) of the needed time. KL.process() computes KL(P||Q) using klPQ() and KL(Q||P) using klQP(). The computations are made once. The second is for free :) ExactBNdistance allows as well to compute in the same time the Hellinger distance ( \(*\sqrt{\sum_i (\sqrt{p_i}-\sqrt{q_i})^2}\)) (Kokolakis and Nanopoulos, 2001).

It may happen that P*ln(P/Q) is not computable (Q=0 and P!=0). In such a case, KL keeps working but trace this error (errorPQ() and errorQP())? *

Warning: This ExactBNdistance should be use only if difficulty()==complexity::CORRECT or at most complexity::DIFFICULT ... snippets :
gum::KL base_kl(net1,net2);

if (base_kl.difficulty()!=KL::HEAVY) {

gum::ExactBNdistance kl(base_kl);

std::cout<<"KL net1||net2 :"<<kl.klPQ()<<std::endl;

} else {

gum::GibbsKL kl(base_kl);

std::cout<<"KL net1||net2 :"<<kl.klPQ()<<std::endl;

}

gum::ExactBNdistance
ExactBNdistance computes exactly the KL divergence betweens 2 BNs.
Definition exactBNdistance.h:89

Definition at line 89 of file exactBNdistance.h.

Constructor & Destructor Documentation

◆ ExactBNdistance() [1/2]

template<typename GUM_SCALAR>

gum::ExactBNdistance< GUM_SCALAR >::ExactBNdistance	(	const IBayesNet< GUM_SCALAR > &	P,
		const IBayesNet< GUM_SCALAR > &	Q )

constructor must give 2 BNs

Exceptions

gum::OperationNotAllowed if the 2 BNs have not the same domainSize or compatible node sets.

Definition at line 58 of file exactBNdistance_tpl.h.

                                                                                   :
      BNdistance< GUM_SCALAR >(P, Q) {
    GUM_CONSTRUCTOR(ExactBNdistance);
  }

References gum::BNdistance< GUM_SCALAR >::BNdistance(), and ExactBNdistance().

Referenced by ExactBNdistance(), ExactBNdistance(), and ~ExactBNdistance().

Here is the call graph for this function:

Here is the caller graph for this function:

◆ ExactBNdistance() [2/2]

template<typename GUM_SCALAR>

gum::ExactBNdistance< GUM_SCALAR >::ExactBNdistance ( const BNdistance< GUM_SCALAR > & kl )

explicit

copy constructor

Definition at line 65 of file exactBNdistance_tpl.h.

                                                                                   :
      BNdistance< GUM_SCALAR >(kl) {
    GUM_CONSTRUCTOR(ExactBNdistance);
  }

References gum::BNdistance< GUM_SCALAR >::BNdistance(), and ExactBNdistance().

Here is the call graph for this function:

◆ ~ExactBNdistance()

template<typename GUM_SCALAR>

gum::ExactBNdistance< GUM_SCALAR >::~ExactBNdistance ( )

virtual

destructor

Definition at line 71 of file exactBNdistance_tpl.h.

                                                  {
    GUM_DESTRUCTOR(ExactBNdistance);
  }

References ExactBNdistance().

Here is the call graph for this function:

Member Function Documentation

◆ _checkCompatibility_()

template<typename GUM_SCALAR>

bool gum::BNdistance< GUM_SCALAR >::_checkCompatibility_ ( ) const

privateinherited

Definition at line 144 of file BNdistance_tpl.h.

                                                            {
    // should not be used
    if (p_.size() != q_.size())
      GUM_ERROR(OperationNotAllowed,
                "BNdistance : the 2 BNs are not compatible (not the same size)")
 
    for (auto node: p_.nodes()) {
      const DiscreteVariable& vp = p_.variable(node);
      try {
        const DiscreteVariable& vq = q_.variableFromName(vp.name());
        if (vp != vq)
          GUM_ERROR(OperationNotAllowed,
                    "BNdistance : the 2 BNs are not compatible "
                    "(not the same variable for the same name : "
                        + vp.toString() + "and " + vq.toString() + ")");
      } catch (NotFound const&) {
        GUM_ERROR(OperationNotAllowed,
                  "BNdistance : the 2 BNs are not compatible (variable : " + vp.name() + ")");
      }
    }
 
    if (std::fabs(p_.log10DomainSize() - q_.log10DomainSize()) > 1e-14) {
      GUM_ERROR(OperationNotAllowed,
                "BNdistance : the 2 BNs are not compatible (not the same domainSize) : p="
                    << p_.log10DomainSize() << " q=" << q_.log10DomainSize() << " => "
                    << p_.log10DomainSize() - q_.log10DomainSize());
    }
 
    return true;
  }

References GUM_ERROR, gum::Variable::name(), p_, q_, and gum::DiscreteVariable::toString().

Referenced by BNdistance().

Here is the call graph for this function:

Here is the caller graph for this function:

◆ bhattacharya()

template<typename GUM_SCALAR>

INLINE double gum::BNdistance< GUM_SCALAR >::bhattacharya ( )

inherited

Returns: Bhattacharya distance (

See also: http://en.wikipedia.org/wiki/Bhattacharya_distance)

Definition at line 109 of file BNdistance_tpl.h.

                                                       {
    process_();
    return bhattacharya_;
  }

References bhattacharya_, and process_().

Here is the call graph for this function:

◆ computeKL_()

template<typename GUM_SCALAR>

void gum::ExactBNdistance< GUM_SCALAR >::computeKL_ ( )

finalprotectedvirtual

Reimplemented from gum::BNdistance< GUM_SCALAR >.

Definition at line 76 of file exactBNdistance_tpl.h.

                                                 {
    klPQ_ = klQP_ = hellinger_ = bhattacharya_ = jsd_ = (GUM_SCALAR)0.0;
    errorPQ_ = errorQP_ = 0;
 
    auto Ip = p_.completeInstantiation();
    auto Iq = q_.completeInstantiation();
 
    // map between p_ variables and q_ variables (using name of vars)
    HashTable< const DiscreteVariable*, const DiscreteVariable* > map;
 
    for (Idx ite = 0; ite < Ip.nbrDim(); ++ite) {
      map.insert(&Ip.variable(ite), &q_.variableFromName(Ip.variable(ite).name()));
    }
    GUM_SCALAR pp, pq, pmid, lpp, lpq, lpmid;
    for (Ip.setFirst(); !Ip.end(); ++Ip) {
      Iq.setValsFrom(map, Ip);
      pp    = p_.jointProbability(Ip);
      pq    = q_.jointProbability(Iq);
      pmid  = (pp + pq) / 2.0;
      lpmid = GUM_LOG2_OR_0(pmid);
      lpp   = GUM_LOG2_OR_0(pp);
      lpq   = GUM_LOG2_OR_0(pq);
 
 
      hellinger_ += std::pow(std::sqrt(pp) - std::sqrt(pq), 2);
      bhattacharya_ += std::sqrt(pp * pq);
 
      if (pp != (GUM_SCALAR)0.0) {
        if (pq != (GUM_SCALAR)0.0) {
          klPQ_ -= pp * (lpq - lpp);   // log2(pq / pp);
        } else {
          errorPQ_++;
        }
      }
 
      if (pq != (GUM_SCALAR)0.0) {
        if (pp != (GUM_SCALAR)0.0) {
          klQP_ -= pq * (lpp - lpq);   // log2(pp / pq);
        } else {
          errorQP_++;
        }
      }
      if (pmid != (GUM_SCALAR)0.0) {
        jsd_ += pp * lpp + pq * lpq
              - (pp + pq) * lpmid;   // pp* log2(pp / pmid) + pq * log2(pq / pmid);
      }
    }
    jsd_ /= 2.0;
    hellinger_    = std::sqrt(hellinger_);
    bhattacharya_ = -std::log(bhattacharya_);
  }

References gum::BNdistance< GUM_SCALAR >::bhattacharya_, gum::BNdistance< GUM_SCALAR >::errorPQ_, gum::BNdistance< GUM_SCALAR >::errorQP_, GUM_LOG2_OR_0, gum::BNdistance< GUM_SCALAR >::hellinger_, gum::HashTable< Key, Val >::insert(), gum::BNdistance< GUM_SCALAR >::jsd_, gum::BNdistance< GUM_SCALAR >::klPQ_, gum::BNdistance< GUM_SCALAR >::klQP_, gum::BNdistance< GUM_SCALAR >::p_, and gum::BNdistance< GUM_SCALAR >::q_.

Here is the call graph for this function:

◆ difficulty()

template<typename GUM_SCALAR>

Complexity gum::BNdistance< GUM_SCALAR >::difficulty ( ) const

inherited

return KL::Complexity::Heavy,KL::Complexity::Difficult,KL::Complexity::Correct depending on the BNs p and q

Definition at line 86 of file BNdistance_tpl.h.

                                                        {
    return _difficulty_;
  }

References _difficulty_.

◆ errorPQ()

template<typename GUM_SCALAR>

INLINE Size gum::BNdistance< GUM_SCALAR >::errorPQ ( )

inherited

Returns: the number of errors while processing divergence KL(P||Q)

Definition at line 121 of file BNdistance_tpl.h.

                                                {
    process_();
    return errorPQ_;
  }

References errorPQ_, and process_().

Here is the call graph for this function:

◆ errorQP()

template<typename GUM_SCALAR>

INLINE Size gum::BNdistance< GUM_SCALAR >::errorQP ( )

inherited

Returns: the number of errors while processing divergence KL(Q||P)

Definition at line 127 of file BNdistance_tpl.h.

                                                {
    process_();
    return errorQP_;
  }

References errorQP_, and process_().

Here is the call graph for this function:

◆ hellinger()

template<typename GUM_SCALAR>

INLINE double gum::BNdistance< GUM_SCALAR >::hellinger ( )

inherited

Returns: hellinger distance (

See also: http://en.wikipedia.org/wiki/Hellinger_distance)

Definition at line 103 of file BNdistance_tpl.h.

                                                    {
    process_();
    return hellinger_;
  }

References hellinger_, and process_().

Here is the call graph for this function:

◆ jsd()

template<typename GUM_SCALAR>

INLINE double gum::BNdistance< GUM_SCALAR >::jsd ( )

inherited

Returns: Jensen-Shannon divergence(

See also: https://en.wikipedia.org/wiki/Jensen%E2%80%93Shannon_divergence)

Definition at line 115 of file BNdistance_tpl.h.

                                              {
    process_();
    return jsd_;
  }

References jsd_, and process_().

Here is the call graph for this function:

◆ klPQ()

template<typename GUM_SCALAR>

INLINE double gum::BNdistance< GUM_SCALAR >::klPQ ( )

inherited

Returns: divergence KL(P||Q)

Definition at line 91 of file BNdistance_tpl.h.

                                               {
    process_();
    return klPQ_;
  }

References klPQ_, and process_().

Here is the call graph for this function:

◆ klQP()

template<typename GUM_SCALAR>

INLINE double gum::BNdistance< GUM_SCALAR >::klQP ( )

inherited

Returns: divergence KL(Q||P)

Definition at line 97 of file BNdistance_tpl.h.

                                               {
    process_();
    return klQP_;
  }

References klQP_, and process_().

Here is the call graph for this function:

◆ p()

template<typename GUM_SCALAR>

INLINE const IBayesNet< GUM_SCALAR > & gum::BNdistance< GUM_SCALAR >::p ( ) const

inherited

Returns: p

Definition at line 133 of file BNdistance_tpl.h.

                                                                          {
    return p_;
  }

References p_.

Referenced by gum::GibbsBNdistance< GUM_SCALAR >::GibbsBNdistance().

Here is the caller graph for this function:

◆ process_()

template<typename GUM_SCALAR>

void gum::BNdistance< GUM_SCALAR >::process_ ( )

protectedinherited

Definition at line 177 of file BNdistance_tpl.h.

                                          {
    if (!_done_) {
      computeKL_();
      _done_ = true;
    }
  }

References _done_, and computeKL_().

Referenced by bhattacharya(), errorPQ(), errorQP(), hellinger(), jsd(), klPQ(), and klQP().

Here is the call graph for this function:

Here is the caller graph for this function:

◆ q()

template<typename GUM_SCALAR>

INLINE const IBayesNet< GUM_SCALAR > & gum::BNdistance< GUM_SCALAR >::q ( ) const

inherited

Returns: q

Definition at line 138 of file BNdistance_tpl.h.

                                                                          {
    return q_;
  }

Definition at line 156 of file BNdistance.h.

Referenced by BNdistance(), BNdistance(), _checkCompatibility_(), gum::ExactBNdistance< GUM_SCALAR >::computeKL_(), gum::GibbsBNdistance< GUM_SCALAR >::computeKL_(), and p().

◆ q_

template<typename GUM_SCALAR>

const IBayesNet< GUM_SCALAR >& gum::BNdistance< GUM_SCALAR >::q_

protectedinherited

Definition at line 157 of file BNdistance.h.

Referenced by BNdistance(), BNdistance(), _checkCompatibility_(), gum::ExactBNdistance< GUM_SCALAR >::computeKL_(), gum::GibbsBNdistance< GUM_SCALAR >::computeKL_(), and q().

The documentation for this class was generated from the following files:

agrum/BN/algorithms/divergence/exactBNdistance.h
agrum/BN/algorithms/divergence/exactBNdistance_tpl.h

Public Member Functions

Protected Member Functions

Protected Attributes

Private Member Functions

Private Attributes

Detailed Description

Constructor & Destructor Documentation

◆ ExactBNdistance() [1/2]

◆ ExactBNdistance() [2/2]

◆ ~ExactBNdistance()

Member Function Documentation

◆ _checkCompatibility_()

◆ bhattacharya()

◆ computeKL_()

◆ difficulty()

◆ errorPQ()

◆ errorQP()

◆ hellinger()

◆ jsd()

◆ klPQ()

◆ klQP()

◆ p()

◆ process_()

◆ q()

Member Data Documentation

◆ _difficulty_

◆ _done_

◆ bhattacharya_

◆ errorPQ_

◆ errorQP_

◆ hellinger_

◆ jsd_

◆ klPQ_

◆ klQP_

◆ p_

◆ q_