FeatureSelection.cpp

浏览该文件的文档.

/*
 * Copyright (c) The Shogun Machine Learning Toolbox
 * Written (w) 2014 Soumyajit De
 * All rights reserved.
 *
 * Redistribution and use in source and binary forms, with or without
 * modification, are permitted provided that the following conditions are met:
 *
 * 1. Redistributions of source code must retain the above copyright notice, this
 *    list of conditions and the following disclaimer.
 * 2. Redistributions in binary form must reproduce the above copyright notice,
 *    this list of conditions and the following disclaimer in the documentation
 *    and/or other materials provided with the distribution.
 *
 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
 * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
 * DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR
 * ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
 * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
 * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
 * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 *
 * The views and conclusions contained in the software and documentation are those
 * of the authors and should not be interpreted as representing official policies,
 * either expressed or implied, of the Shogun Development Team.
 */

#include <shogun/labels/Labels.h>
#include <shogun/features/Features.h>
#include <shogun/features/DenseFeatures.h>
#include <shogun/features/SparseFeatures.h>
#include <shogun/features/SubsetStack.h>
#include <shogun/preprocessor/FeatureSelection.h>
#include <shogun/mathematics/Math.h>

namespace shogun
{

template <class ST>
CFeatureSelection<ST>::CFeatureSelection() : CPreprocessor()
{
    initialize_parameters();
}

template <class ST>
void CFeatureSelection<ST>::initialize_parameters()
{
    SG_ADD(&m_target_dim, "target_dim", "target dimension",
            MS_NOT_AVAILABLE);
    SG_ADD((machine_int_t*)&m_algorithm, "algorithm",
            "the feature selectiona algorithm", MS_NOT_AVAILABLE);
    SG_ADD((machine_int_t*)&m_policy, "policy", "feature removal policy",
            MS_NOT_AVAILABLE);
    SG_ADD(&m_num_remove, "num_remove", "number or percentage of features to "
            "be removed", MS_NOT_AVAILABLE);
    SG_ADD((CSGObject**)&m_labels, "labels",
            "the class labels for the features", MS_NOT_AVAILABLE);
    SG_ADD((CSGObject**)&m_subset, "subset",
            "indices of selected features", MS_NOT_AVAILABLE);

    m_target_dim=0;
    m_algorithm=BACKWARD_ELIMINATION;
    m_policy=N_LARGEST;
    m_num_remove=1;
    m_labels=NULL;
    m_subset=new CSubsetStack();
}

template <class ST>
CFeatureSelection<ST>::~CFeatureSelection()
{
    SG_UNREF(m_labels);
    SG_UNREF(m_subset);
}


template <class ST>
void CFeatureSelection<ST>::cleanup()
{
    m_subset->remove_all_subsets();
}

template <class ST>
CFeatures* CFeatureSelection<ST>::apply_backward_elimination(CFeatures* features)
{
    SG_DEBUG("Entering!\n");

    // precompute whenever appropriate for performing the rest of the tasks
    precompute();

    // NULL check for features is handled in get_num_features
    index_t num_features=get_num_features(features);
    SG_DEBUG("Initial number of features %d!\n", num_features);

    // the main loop
    while (num_features>m_target_dim)
    {
        // tune the measurement parameters whenever necessary based on current
        // features
        adapt_params(features);

        // compute the measures for each of the current dimensions
        SGVector<float64_t> measures(num_features);
        for (index_t i=0; i<num_features; ++i)
            measures[i]=compute_measures(features, i);

        if (io->get_loglevel()==MSG_DEBUG || io->get_loglevel()==MSG_GCDEBUG)
            measures.display_vector("measures");

        // rank the measures
        SGVector<index_t> argsorted=CMath::argsort(measures);

        if (io->get_loglevel()==MSG_DEBUG || io->get_loglevel()==MSG_GCDEBUG)
            argsorted.display_vector("argsorted");

        // make sure that we don't end up with lesser feats than target dim
        index_t to_remove;
        if (m_policy==N_SMALLEST || m_policy==N_LARGEST)
            to_remove=m_num_remove;
        else
            to_remove=num_features*m_num_remove*0.01;

        index_t can_remove=num_features-m_target_dim;

        // if policy is to remove N feats corresponding to smallest/largest
        // measures, we just replace N with can_remove. if policy is to remove
        // N% feats, then we change the policy temporarily and remove a fixed
        // can_remove number of feats instead
        index_t orig_remove=m_num_remove;
        EFeatureRemovalPolicy orig_policy=m_policy;

        if (to_remove>can_remove)
        {
            m_num_remove=can_remove;
            SG_DEBUG("Can only remove %d features in this iteration!\n",
                    can_remove);

            if (m_policy==PERCENTILE_SMALLEST)
                m_policy=N_SMALLEST;
            else if (m_policy==PERCENTILE_LARGEST)
                m_policy=N_LARGEST;
        }

        // remove appropriate number of features based on the measures and the
        // removal policy. this internally update the subset for selected
        // features as well
        features=remove_feats(features, argsorted);

        // restore original removal policy and numbers if necessary for the
        // sake of consistency
        if (to_remove>can_remove)
        {
            m_policy=orig_policy;
            m_num_remove=orig_remove;
        }

        // update the number of features
        num_features=get_num_features(features);
        SG_DEBUG("Current number of features %d!\n", num_features);
    }

    // sanity check
    ASSERT(m_subset->get_size()==m_target_dim);

    SG_DEBUG("Leaving!\n");
    return features;
}

template <class ST>
CFeatures* CFeatureSelection<ST>::apply(CFeatures* features)
{
    SG_DEBUG("Entering!\n");

    // remove previously computed feature subsets
    m_subset->remove_all_subsets();

    // sanity checks
    REQUIRE(features, "Features cannot be NULL!\n");
    REQUIRE(features->get_num_vectors()>0,
            "Number of feature vectors has to be positive!\n");
    REQUIRE(m_target_dim>0, "Target dimension (%d) has to be positive! Set "
            "a higher number via set_target_dim().\n", m_target_dim);

    index_t num_features=get_num_features(features);
    REQUIRE(num_features>0, "Invalid number of features (%d)! Most likely "
            "feature selection cannot be performed for %s!\n",
            num_features, features->get_name());
    REQUIRE(num_features>m_target_dim,
            "Number of original features (dimensions of the feature vectors) "
            "(%d) has to be greater that the target dimension (%d)!\n",
            num_features, m_target_dim);

    // this method makes a deep copy of the feature object and performs
    // feature selection on it. This is already SG_REF'ed because of the
    // implementation of clone()
    CFeatures* feats_copy=(CFeatures*)features->clone();

    switch (m_algorithm)
    {
        case BACKWARD_ELIMINATION:
            return apply_backward_elimination(feats_copy);
        default:
            SG_ERROR("Specified algorithm not yet supported!\n");
            return features;
    }

    SG_DEBUG("Leaving!\n");
}

template <class ST>
void CFeatureSelection<ST>::precompute()
{
}

template <class ST>
void CFeatureSelection<ST>::adapt_params(CFeatures* features)
{
}

template <class ST>
SGVector<index_t> CFeatureSelection<ST>::get_selected_feats()
{
    ASSERT(m_subset);

    SGVector<index_t> inds;
    if (m_subset->has_subsets())
    {
        inds=SGVector<index_t>(m_subset->get_size());
        for (index_t i=0; i<inds.vlen; ++i)
            inds[i]=m_subset->subset_idx_conversion(i);
        CMath::qsort(inds);
    }

    return inds;
}

template <class ST>
index_t CFeatureSelection<ST>::get_num_features(CFeatures* features) const
{
    REQUIRE(features, "Features not initialized!\n");

    EFeatureClass f_class=features->get_feature_class();

    switch (f_class)
    {
        case C_DENSE:
        {
            CDenseFeatures<ST>* d_feats=dynamic_cast<CDenseFeatures<ST>*>(features);
            REQUIRE(d_feats, "Type mismatch for dense features!\n");
            return d_feats->get_num_features();
        }
        case C_SPARSE:
        {
            CSparseFeatures<ST>* s_feats=dynamic_cast<CSparseFeatures<ST>*>(features);
            REQUIRE(s_feats, "Type mismatch for sparse features!\n");
            return s_feats->get_num_features();
        }
        default:
            SG_ERROR("Number of features not available for %s!\n",
                    features->get_name());
            break;
    }

    return 0;
}

template <class ST>
void CFeatureSelection<ST>::set_target_dim(index_t target_dim)
{
    m_target_dim=target_dim;
}

template <class ST>
index_t CFeatureSelection<ST>::get_target_dim() const
{
    return m_target_dim;
}

template <class ST>
EFeatureSelectionAlgorithm CFeatureSelection<ST>::get_algorithm() const
{
    return m_algorithm;
}

template <class ST>
EFeatureRemovalPolicy CFeatureSelection<ST>::get_policy() const
{
    return m_policy;
}

template <class ST>
void CFeatureSelection<ST>::set_num_remove(index_t num_remove)
{
    m_num_remove=num_remove;
}

template <class ST>
index_t CFeatureSelection<ST>::get_num_remove() const
{
    return m_num_remove;
}

template <class ST>
void CFeatureSelection<ST>::set_labels(CLabels* labels)
{
    SG_REF(labels);
    SG_UNREF(m_labels);
    m_labels=labels;
}

template <class ST>
CLabels* CFeatureSelection<ST>::get_labels() const
{
    SG_REF(m_labels);
    return m_labels;
}

template <class ST>
EFeatureClass CFeatureSelection<ST>::get_feature_class()
{
    return C_ANY;
}

template <class ST>
EPreprocessorType CFeatureSelection<ST>::get_type() const
{
    return P_UNKNOWN;
}

template<>
EFeatureType CFeatureSelection<floatmax_t>::get_feature_type()
{
    return F_LONGREAL;
}

template<>
EFeatureType CFeatureSelection<float64_t>::get_feature_type()
{
    return F_DREAL;
}

template<>
EFeatureType CFeatureSelection<float32_t>::get_feature_type()
{
    return F_SHORTREAL;
}

template<>
EFeatureType CFeatureSelection<int16_t>::get_feature_type()
{
    return F_SHORT;
}

template<>
EFeatureType CFeatureSelection<uint16_t>::get_feature_type()
{
    return F_WORD;
}

template<>
EFeatureType CFeatureSelection<char>::get_feature_type()
{
    return F_CHAR;
}

template<>
EFeatureType CFeatureSelection<int8_t>::get_feature_type()
{
    return F_CHAR;
}

template<>
EFeatureType CFeatureSelection<uint8_t>::get_feature_type()
{
    return F_BYTE;
}

template<>
EFeatureType CFeatureSelection<int32_t>::get_feature_type()
{
    return F_INT;
}

template<>
EFeatureType CFeatureSelection<uint32_t>::get_feature_type()
{
    return F_UINT;
}

template<>
EFeatureType CFeatureSelection<int64_t>::get_feature_type()
{
    return F_LONG;
}

template<>
EFeatureType CFeatureSelection<uint64_t>::get_feature_type()
{
    return F_ULONG;
}

template<>
EFeatureType CFeatureSelection<bool>::get_feature_type()
{
    return F_BOOL;
}

template class CFeatureSelection<bool>;
template class CFeatureSelection<char>;
template class CFeatureSelection<int8_t>;
template class CFeatureSelection<uint8_t>;
template class CFeatureSelection<int16_t>;
template class CFeatureSelection<uint16_t>;
template class CFeatureSelection<int32_t>;
template class CFeatureSelection<uint32_t>;
template class CFeatureSelection<int64_t>;
template class CFeatureSelection<uint64_t>;
template class CFeatureSelection<float32_t>;
template class CFeatureSelection<float64_t>;
template class CFeatureSelection<floatmax_t>;

}