scoreBIC.cpp

Go to the documentation of this file.

/****************************************************************************
 *   This file is part of the aGrUM/pyAgrum library.                        *
 *                                                                          *
 *   Copyright (c) 2005-2026 by                                             *
 *       - Pierre-Henri WUILLEMIN(_at_LIP6)                                 *
 *       - Christophe GONZALES(_at_AMU)                                     *
 *                                                                          *
 *   The aGrUM/pyAgrum library is free software; you can redistribute it    *
 *   and/or modify it under the terms of either :                           *
 *                                                                          *
 *    - the GNU Lesser General Public License as published by               *
 *      the Free Software Foundation, either version 3 of the License,      *
 *      or (at your option) any later version,                              *
 *    - the MIT license (MIT),                                              *
 *    - or both in dual license, as here.                                   *
 *                                                                          *
 *   (see https://agrum.gitlab.io/articles/dual-licenses-lgplv3mit.html)    *
 *                                                                          *
 *   This aGrUM/pyAgrum library is distributed in the hope that it will be  *
 *   useful, but WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED,          *
 *   INCLUDING BUT NOT LIMITED TO THE WARRANTIES MERCHANTABILITY or FITNESS *
 *   FOR A PARTICULAR PURPOSE  AND NONINFRINGEMENT. IN NO EVENT SHALL THE   *
 *   AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER *
 *   LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,        *
 *   ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR  *
 *   OTHER DEALINGS IN THE SOFTWARE.                                        *
 *                                                                          *
 *   See LICENCES for more details.                                         *
 *                                                                          *
 *   SPDX-FileCopyrightText: Copyright 2005-2026                            *
 *       - Pierre-Henri WUILLEMIN(_at_LIP6)                                 *
 *       - Christophe GONZALES(_at_AMU)                                     *
 *   SPDX-License-Identifier: LGPL-3.0-or-later OR MIT                      *
 *                                                                          *
 *   Contact  : info_at_agrum_dot_org                                       *
 *   homepage : http://agrum.gitlab.io                                      *
 *   gitlab   : https://gitlab.com/agrumery/agrum                           *
 *                                                                          *
 ****************************************************************************/
 

 
#include <agrum/BN/learning/scores/scoreBIC.h>
 
#ifndef DOXYGEN_SHOULD_SKIP_THIS

#  ifdef GUM_NO_INLINE
#    include <agrum/BN/learning/scores/scoreBIC_inl.h>
#  endif /* GUM_NO_INLINE */
 
namespace gum {
 
  namespace learning {
 
    // Constructors and destructor are defined out-of-line (not INLINE) on
    // purpose: see the comment in score.cpp for the MSVC LNK2005 rationale.

    ScoreBIC::ScoreBIC(const DBRowGeneratorParser&                                 parser,
                       const Prior&                                                prior,
                       const std::vector< std::pair< std::size_t, std::size_t > >& ranges,
                       const Bijection< NodeId, std::size_t >&                     nodeId2columns) :
        Score(parser, prior, ranges, nodeId2columns),
        _internal_prior_(parser.database(), nodeId2columns) {
      GUM_CONSTRUCTOR(ScoreBIC);
    }

    ScoreBIC::ScoreBIC(const DBRowGeneratorParser&             parser,
                       const Prior&                            prior,
                       const Bijection< NodeId, std::size_t >& nodeId2columns) :
        Score(parser, prior, nodeId2columns), _internal_prior_(parser.database(), nodeId2columns) {
      GUM_CONSTRUCTOR(ScoreBIC);
    }

    ScoreBIC::ScoreBIC(const ScoreBIC& from) :
        Score(from), _internal_prior_(from._internal_prior_) {
      GUM_CONS_CPY(ScoreBIC);
    }

    ScoreBIC::ScoreBIC(ScoreBIC&& from) :
        Score(std::move(from)), _internal_prior_(std::move(from._internal_prior_)) {
      GUM_CONS_MOV(ScoreBIC);
    }

    ScoreBIC::~ScoreBIC() { GUM_DESTRUCTOR(ScoreBIC); }

    ScoreBIC& ScoreBIC::operator=(const ScoreBIC& from) {
      if (this != &from) {
        Score::operator=(from);
        _internal_prior_ = from._internal_prior_;
      }
      return *this;
    }

    ScoreBIC& ScoreBIC::operator=(ScoreBIC&& from) {
      if (this != &from) {
        Score::operator=(std::move(from));
        _internal_prior_ = std::move(from._internal_prior_);
      }
      return *this;
    }

    std::string ScoreBIC::isPriorCompatible(PriorType prior_type, double weight) {
      // check that the prior is compatible with the score
      if ((prior_type == PriorType::DirichletPriorType)
          || (prior_type == PriorType::SmoothingPriorType)
          || (prior_type == PriorType::NoPriorType)) {
        return "";
      }
 
      // prior types unsupported by the type checker
      return std::format("The prior '{}' is not yet compatible with the score 'BIC'.",
                         priorTypeToString(prior_type));
    }

    double ScoreBIC::score_(const IdCondSet& idset) {
      // get the counts for all the nodes in the idset and add the prior
      std::vector< double > N_ijk(this->counter_.counts(idset, true));
      const bool            informative_external_prior = this->prior_->isInformative();
      if (informative_external_prior) this->prior_->addJointPseudoCount(idset, N_ijk);
      const std::size_t all_size = N_ijk.size();
 
      // here, we distinguish idsets with conditioning nodes from those
      // without conditioning nodes
      if (idset.hasConditioningSet()) {
        // get the counts for the conditioning nodes
        std::vector< double > N_ij(this->marginalize_(idset[0], N_ijk));
        const std::size_t     conditioning_size = N_ij.size();
 
        // initialize the score: this should be the penalty of the BIC score,
        // i.e., -(ri-1 ) * qi * .5 * log ( N + N' )
        const std::size_t target_domsize = all_size / conditioning_size;
        const double      penalty = conditioning_size * double(target_domsize - std::size_t(1));
 
        // compute the score: it remains to compute the log likelihood, i.e.,
        // sum_k=1^r_i sum_j=1^q_i N_ijk log (N_ijk / N_ij), which is also
        // equivalent to:
        // sum_j=1^q_i sum_k=1^r_i N_ijk log N_ijk - sum_j=1^q_i N_ij log N_ij
        double score = 0.0;
        for (const auto n_ijk: N_ijk) {
          if (n_ijk) { score += n_ijk * std::log(n_ijk); }
        }
        double N = 0;
        for (const auto n_ij: N_ij) {
          if (n_ij) {
            score -= n_ij * std::log(n_ij);
            N += n_ij;
          }
        }
 
        // finally, remove the penalty
        score -= penalty * std::log(N) * 0.5;
 
        // divide by log(2), since the log likelihood uses log_2
        score *= this->one_log2_;
 
        return score;
      } else {
        // here, there are no conditioning nodes
 
        // initialize the score: this should be the penalty of the BIC score,
        // i.e., -(ri-1 )
        const double penalty = double(all_size - std::size_t(1));
 
        // compute the score: it remains to compute the log likelihood, i.e.,
        // sum_k=1^r_i N_ijk log (N_ijk / N), which is also
        // equivalent to:
        // sum_j=1^q_i sum_k=1^r_i N_ijk log N_ijk - N log N
        double N     = 0.0;
        double score = 0.0;
        for (const auto n_ijk: N_ijk) {
          if (n_ijk) {
            score += n_ijk * std::log(n_ijk);
            N += n_ijk;
          }
        }
        score -= N * std::log(N);
 
        // finally, remove the penalty
        score -= penalty * std::log(N) * 0.5;
 
        // divide by log(2), since the log likelihood uses log_2
        score *= this->one_log2_;
 
        return score;
      }
    }

    double ScoreBIC::N(const IdCondSet& idset) {
      // get the counts for all the nodes in the idset and add the prior
      std::vector< double > N_ijk(this->counter_.counts(idset, true));
      if (this->prior_->isInformative()) this->prior_->addJointPseudoCount(idset, N_ijk);
 
      double N = 0;
      for (const auto n_ijk: N_ijk) {
        N += n_ijk;
      }
 
      return N;
    }
 
  } /* namespace learning */
 
} /* namespace gum */
 
#endif /* DOXYGEN_SHOULD_SKIP_THIS */