GaussianNaiveBayes.cpp

Go to the documentation of this file.

/*
 * This program is free software; you can redistribute it and/or modify
 * it under the terms of the GNU General Public License as published by
 * the Free Software Foundation; either version 3 of the License, or
 * (at your option) any later version.
 *
 * Written (W) 2011 Sergey Lisitsyn
 * Copyright (C) 2011 Berlin Institute of Technology and Max-Planck-Society
 */

#include <shogun/multiclass/GaussianNaiveBayes.h>
#include <shogun/features/Features.h>
#include <shogun/labels/Labels.h>
#include <shogun/labels/RegressionLabels.h>
#include <shogun/labels/MulticlassLabels.h>
#include <shogun/mathematics/Math.h>
#include <shogun/lib/Signal.h>

using namespace shogun;

CGaussianNaiveBayes::CGaussianNaiveBayes() : CNativeMulticlassMachine(), m_features(NULL),
    m_min_label(0), m_num_classes(0), m_dim(0), m_means(), m_variances(),
    m_label_prob(), m_rates()
{

};

CGaussianNaiveBayes::CGaussianNaiveBayes(CFeatures* train_examples,
    CLabels* train_labels) : CNativeMulticlassMachine(), m_features(NULL),
    m_min_label(0), m_num_classes(0), m_dim(0), m_means(),
    m_variances(), m_label_prob(), m_rates()
{
    ASSERT(train_examples->get_num_vectors() == train_labels->get_num_labels())
    set_labels(train_labels);

    if (!train_examples->has_property(FP_DOT))
        SG_ERROR("Specified features are not of type CDotFeatures\n")

    set_features((CDotFeatures*)train_examples);
};

CGaussianNaiveBayes::~CGaussianNaiveBayes()
{
    SG_UNREF(m_features);
};

CFeatures* CGaussianNaiveBayes::get_features()
{
    SG_REF(m_features);
    return m_features;
}

void CGaussianNaiveBayes::set_features(CFeatures* features)
{
    if (!features->has_property(FP_DOT))
        SG_ERROR("Specified features are not of type CDotFeatures\n")

    SG_REF(features);
    SG_UNREF(m_features);
    m_features = (CDotFeatures*)features;
}

bool CGaussianNaiveBayes::train_machine(CFeatures* data)
{
    // init features with data if necessary and assure type is correct
    if (data)
    {
        if (!data->has_property(FP_DOT))
                SG_ERROR("Specified features are not of type CDotFeatures\n")
        set_features((CDotFeatures*) data);
    }

    // get int labels to train_labels and check length equality
    ASSERT(m_labels)
    ASSERT(m_labels->get_label_type() == LT_MULTICLASS)
    SGVector<int32_t> train_labels = ((CMulticlassLabels*) m_labels)->get_int_labels();
    ASSERT(m_features->get_num_vectors()==train_labels.vlen)

    // init min_label, max_label and loop variables
    int32_t min_label = train_labels.vector[0];
    int32_t max_label = train_labels.vector[0];
    int i,j;

    // find minimal and maximal label
    for (i=1; i<train_labels.vlen; i++)
    {
        min_label = CMath::min(min_label, train_labels.vector[i]);
        max_label = CMath::max(max_label, train_labels.vector[i]);
    }

    // subtract minimal label from all labels
    for (i=0; i<train_labels.vlen; i++)
        train_labels.vector[i]-= min_label;

    // get number of classes, minimal label and dimensionality
    m_num_classes = max_label-min_label+1;
    m_min_label = min_label;
    m_dim = m_features->get_dim_feature_space();

    // allocate memory for distributions' parameters and a priori probability
    m_means=SGMatrix<float64_t>(m_dim,m_num_classes);
    m_variances=SGMatrix<float64_t>(m_dim, m_num_classes);
    m_label_prob=SGVector<float64_t>(m_num_classes);

    // allocate memory for label rates
    m_rates=SGVector<float64_t>(m_num_classes);

    // make arrays filled by zeros before using
    m_means.zero();
    m_variances.zero();
    m_label_prob.zero();
    m_rates.zero();

    // number of iterations in all cycles
    int32_t max_progress = 2 * train_labels.vlen + 2 * m_num_classes;
    
    // current progress
    int32_t progress = 0;   
    SG_PROGRESS(progress, 0, max_progress)

    // get sum of features among labels
    for (i=0; i<train_labels.vlen; i++)
    {
        SGVector<float64_t> fea = m_features->get_computed_dot_feature_vector(i);
        for (j=0; j<m_dim; j++)
            m_means(j, train_labels.vector[i]) += fea.vector[j];

        m_label_prob.vector[train_labels.vector[i]]+=1.0;

        progress++;
        SG_PROGRESS(progress, 0, max_progress)
    }

    // get means of features of labels
    for (i=0; i<m_num_classes; i++)
    {
        for (j=0; j<m_dim; j++)
            m_means(j, i) /= m_label_prob.vector[i];

        progress++;
        SG_PROGRESS(progress, 0, max_progress)
    }

    // compute squared residuals with means available
    for (i=0; i<train_labels.vlen; i++)
    {
        SGVector<float64_t> fea = m_features->get_computed_dot_feature_vector(i);
        for (j=0; j<m_dim; j++)
        {
            m_variances(j, train_labels.vector[i]) += 
                CMath::sq(fea[j]-m_means(j, train_labels.vector[i]));
        }

        progress++;
        SG_PROGRESS(progress, 0, max_progress)
    }   

    // get variance of features of labels
    for (i=0; i<m_num_classes; i++)
    {
        for (j=0; j<m_dim; j++)
            m_variances(j, i) /= m_label_prob.vector[i] > 1 ? m_label_prob.vector[i]-1 : 1;
        
        // get a priori probabilities of labels
        m_label_prob.vector[i]/= m_num_classes;

        progress++;
        SG_PROGRESS(progress, 0, max_progress)
    }
    SG_DONE()

    return true;
}

CMulticlassLabels* CGaussianNaiveBayes::apply_multiclass(CFeatures* data)
{
    if (data)
        set_features(data);

    ASSERT(m_features)

    // init number of vectors
    int32_t num_vectors = m_features->get_num_vectors();

    // init result labels
    CMulticlassLabels* result = new CMulticlassLabels(num_vectors);

    // classify each example of data
    SG_PROGRESS(0, 0, num_vectors)
    for (int i = 0; i < num_vectors; i++)
    {
        result->set_label(i,apply_one(i));
        SG_PROGRESS(i + 1, 0, num_vectors)
    }
    SG_DONE()
    return result;
};

float64_t CGaussianNaiveBayes::apply_one(int32_t idx)
{
    // get [idx] feature vector
    SGVector<float64_t> feature_vector = m_features->get_computed_dot_feature_vector(idx);

    // init loop variables
    int i,k;

    // rate all labels
    for (i=0; i<m_num_classes; i++)
    {
        // set rate to 0.0 if a priori probability is 0.0 and continue
        if (m_label_prob.vector[i]==0.0)
        {
            m_rates.vector[i] = 0.0;
            continue;
        }
        else
            m_rates.vector[i] = CMath::log(m_label_prob.vector[i]);
    
        // product all conditional gaussian probabilities
        for (k=0; k<m_dim; k++)
            if (m_variances(k,i)!=0.0)
                m_rates.vector[i]+= CMath::log(0.39894228/CMath::sqrt(m_variances(k, i))) - 
                    0.5*CMath::sq(feature_vector.vector[k]-m_means(k, i))/(m_variances(k, i));
    }

    // find label with maximum rate
    int32_t max_label_idx = 0;

    for (i=0; i<m_num_classes; i++)
    {
        if (m_rates.vector[i]>m_rates.vector[max_label_idx])
            max_label_idx = i;
    }

    return max_label_idx+m_min_label;
};