StudentsTLikelihood.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) 2013 Roman Votyakov
 * Copyright (C) 2012 Jacob Walker
 * Copyright (C) 2013 Roman Votyakov
 *
 * Adapted from the GPML toolbox, specifically likT.m
 * http://www.gaussianprocess.org/gpml/code/matlab/doc/
 */

#include <shogun/machine/gp/StudentsTLikelihood.h>

#ifdef HAVE_EIGEN3

#include <shogun/mathematics/Function.h>
#include <shogun/mathematics/Integration.h>
#include <shogun/labels/RegressionLabels.h>
#include <shogun/mathematics/Statistics.h>
#include <shogun/mathematics/Math.h>
#include <shogun/mathematics/eigen3.h>

using namespace shogun;
using namespace Eigen;

namespace shogun
{

#ifndef DOXYGEN_SHOULD_SKIP_THIS

class CNormalPDF : public CFunction
{
public:
    CNormalPDF()
    {
        m_mu=0.0;
        m_sigma=1.0;
    }

    CNormalPDF(float64_t mu, float64_t sigma)
    {
        m_mu=mu;
        m_sigma=sigma;
    }

    void set_mu(float64_t mu) { m_mu=mu; }

    void set_sigma(float64_t sigma) { m_sigma=sigma; }

    virtual float64_t operator() (float64_t x)
    {
        return (1.0/(CMath::sqrt(2*CMath::PI)*m_sigma))*
            CMath::exp(-CMath::sq(x-m_mu)/(2.0*CMath::sq(m_sigma)));
    }

private:
    /* mean */
    float64_t m_mu;

    /* standard deviation */
    float64_t m_sigma;
};

class CStudentsTPDF : public CFunction
{
public:
    CStudentsTPDF()
    {
        m_sigma=1.0;
        m_mu=0.0;
        m_nu=3.0;
    }

    CStudentsTPDF(float64_t sigma, float64_t nu, float64_t mu)
    {
        m_sigma=sigma;
        m_mu=mu;
        m_nu=nu;
    }

    void set_sigma(float64_t sigma) { m_sigma=sigma; }

    void set_mu(float64_t mu) { m_mu=mu; }

    void set_nu(float64_t nu) { m_nu=nu; }

    virtual float64_t operator() (float64_t x)
    {
        float64_t lZ=CStatistics::lgamma((m_nu+1.0)/2.0)-CStatistics::lgamma(m_nu/2.0)-
            CMath::log(m_nu*CMath::PI*CMath::sq(m_sigma))/2.0;
        return CMath::exp(lZ-((m_nu+1.0)/2.0)*CMath::log(1.0+CMath::sq(x-m_mu)/
            (m_nu*CMath::sq(m_sigma))));
    }

private:
    float64_t m_sigma;

    float64_t m_mu;

    float64_t m_nu;
};

class CProductFunction : public CFunction
{
public:
    CProductFunction(CFunction* f, CFunction* g)
    {
        SG_REF(f);
        SG_REF(g);
        m_f=f;
        m_g=g;
    }

    virtual ~CProductFunction()
    {
        SG_UNREF(m_f);
        SG_UNREF(m_g);
    }

    virtual float64_t operator() (float64_t x)
    {
        return (*m_f)(x)*(*m_g)(x);
    }

private:
    CFunction* m_f;
    CFunction* m_g;
};

class CLinearFunction : public CFunction
{
public:
    CLinearFunction() { }

    virtual ~CLinearFunction() { }

    virtual float64_t operator() (float64_t x)
    {
        return x;
    }
};

class CQuadraticFunction : public CFunction
{
public:
    CQuadraticFunction() { }

    virtual ~CQuadraticFunction() { }

    virtual float64_t operator() (float64_t x)
    {
        return CMath::sq(x);
    }
};

#endif /* DOXYGEN_SHOULD_SKIP_THIS */

CStudentsTLikelihood::CStudentsTLikelihood() : CLikelihoodModel()
{
    init();
}

CStudentsTLikelihood::CStudentsTLikelihood(float64_t sigma, float64_t df)
        : CLikelihoodModel()
{
    REQUIRE(sigma>0.0, "Scale parameter must be greater than zero\n")
    REQUIRE(df>1.0, "Number of degrees of freedom must be greater than one\n")

    init();
    m_sigma=sigma;
    m_df=df;
}

void CStudentsTLikelihood::init()
{
    m_sigma=1.0;
    m_df=3.0;
    SG_ADD(&m_df, "df", "Degrees of freedom", MS_AVAILABLE, GRADIENT_AVAILABLE);
    SG_ADD(&m_sigma, "sigma", "Scale parameter", MS_AVAILABLE, GRADIENT_AVAILABLE);
}

CStudentsTLikelihood::~CStudentsTLikelihood()
{
}

CStudentsTLikelihood* CStudentsTLikelihood::obtain_from_generic(
        CLikelihoodModel* lik)
{
    ASSERT(lik!=NULL);

    if (lik->get_model_type()!=LT_STUDENTST)
        SG_SERROR("Provided likelihood is not of type CStudentsTLikelihood!\n")

    SG_REF(lik);
    return (CStudentsTLikelihood*)lik;
}

SGVector<float64_t> CStudentsTLikelihood::get_predictive_means(
        SGVector<float64_t> mu, SGVector<float64_t> s2, const CLabels* lab) const
{
    return SGVector<float64_t>(mu);
}

SGVector<float64_t> CStudentsTLikelihood::get_predictive_variances(
        SGVector<float64_t> mu, SGVector<float64_t> s2, const CLabels* lab) const
{
    SGVector<float64_t> result(s2);
    Map<VectorXd> eigen_result(result.vector, result.vlen);

    if (m_df<2.0)
        eigen_result=CMath::INFTY*VectorXd::Ones(result.vlen);
    else
    {
        eigen_result+=CMath::sq(m_sigma)*m_df/(m_df-2.0)*
            VectorXd::Ones(result.vlen);
    }

    return result;
}

SGVector<float64_t> CStudentsTLikelihood::get_log_probability_f(const CLabels* lab,
        SGVector<float64_t> func) const
{
    // check the parameters
    REQUIRE(lab, "Labels are required (lab should not be NULL)\n")
    REQUIRE(lab->get_label_type()==LT_REGRESSION,
            "Labels must be type of CRegressionLabels\n")
    REQUIRE(lab->get_num_labels()==func.vlen, "Number of labels must match "
            "length of the function vector\n")

    Map<VectorXd> eigen_f(func.vector, func.vlen);

    SGVector<float64_t> r(func.vlen);
    Map<VectorXd> eigen_r(r.vector, r.vlen);

    SGVector<float64_t> y=((CRegressionLabels*)lab)->get_labels();
    Map<VectorXd> eigen_y(y.vector, y.vlen);

    // compute lZ=log(gamma(df/2+1/2))-log(gamma(df/2))-log(df*pi*sigma^2)/2
    VectorXd eigen_lZ=(CStatistics::lgamma(m_df/2.0+0.5)-
        CStatistics::lgamma(m_df/2.0)-log(m_df*CMath::PI*CMath::sq(m_sigma))/2.0)*
        VectorXd::Ones(r.vlen);

    // compute log probability: lp=lZ-(df+1)*log(1+(y-f).^2./(df*sigma^2))/2
    eigen_r=eigen_y-eigen_f;
    eigen_r=eigen_r.cwiseProduct(eigen_r)/(m_df*CMath::sq(m_sigma));
    eigen_r=eigen_lZ-(m_df+1)*
        (eigen_r+VectorXd::Ones(r.vlen)).array().log().matrix()/2.0;

    return r;
}

SGVector<float64_t> CStudentsTLikelihood::get_log_probability_derivative_f(
        const CLabels* lab, SGVector<float64_t> func, index_t i) const
{
    // check the parameters
    REQUIRE(lab, "Labels are required (lab should not be NULL)\n")
    REQUIRE(lab->get_label_type()==LT_REGRESSION,
            "Labels must be type of CRegressionLabels\n")
    REQUIRE(lab->get_num_labels()==func.vlen, "Number of labels must match "
            "length of the function vector\n")
    REQUIRE(i>=1 && i<=3, "Index for derivative should be 1, 2 or 3\n")

    Map<VectorXd> eigen_f(func.vector, func.vlen);

    SGVector<float64_t> r(func.vlen);
    Map<VectorXd> eigen_r(r.vector, r.vlen);

    SGVector<float64_t> y=((CRegressionLabels*)lab)->get_labels();
    Map<VectorXd> eigen_y(y.vector, y.vlen);

    // compute r=y-f, r2=r.^2
    eigen_r=eigen_y-eigen_f;
    VectorXd eigen_r2=eigen_r.cwiseProduct(eigen_r);

    // compute a=(y-f).^2+df*sigma^2
    VectorXd a=eigen_r2+VectorXd::Ones(r.vlen)*m_df*CMath::sq(m_sigma);

    if (i==1)
    {
        // compute first derivative of log probability wrt f:
        // dlp=(df+1)*(y-f)./a
        eigen_r=(m_df+1)*eigen_r.cwiseQuotient(a);
    }
    else if (i==2)
    {
        // compute second derivative of log probability wrt f:
        // d2lp=(df+1)*((y-f)^2-df*sigma^2)./a.^2
        VectorXd b=eigen_r2-VectorXd::Ones(r.vlen)*m_df*CMath::sq(m_sigma);

        eigen_r=(m_df+1)*b.cwiseQuotient(a.cwiseProduct(a));
    }
    else if (i==3)
    {
        // compute third derivative of log probability wrt f:
        // d3lp=(f+1)*2*(y-f).*((y-f)^2-3*f*sigma^2)./a.^3
        VectorXd c=eigen_r2-VectorXd::Ones(r.vlen)*3*m_df*CMath::sq(m_sigma);
        VectorXd a2=a.cwiseProduct(a);

        eigen_r=(m_df+1)*2*eigen_r.cwiseProduct(c).cwiseQuotient(
            a2.cwiseProduct(a));
    }

    return r;
}

SGVector<float64_t> CStudentsTLikelihood::get_first_derivative(const CLabels* lab,
        SGVector<float64_t> func, const TParameter* param) const
{
    // check the parameters
    REQUIRE(lab, "Labels are required (lab should not be NULL)\n")
    REQUIRE(lab->get_label_type()==LT_REGRESSION,
            "Labels must be type of CRegressionLabels\n")
    REQUIRE(lab->get_num_labels()==func.vlen, "Number of labels must match "
            "length of the function vector\n")

    SGVector<float64_t> r(func.vlen);
    Map<VectorXd> eigen_r(r.vector, r.vlen);

    Map<VectorXd> eigen_f(func.vector, func.vlen);

    SGVector<float64_t> y=((CRegressionLabels*)lab)->get_labels();
    Map<VectorXd> eigen_y(y.vector, y.vlen);

    // compute r=y-f and r2=(y-f).^2
    eigen_r=eigen_y-eigen_f;
    VectorXd eigen_r2=eigen_r.cwiseProduct(eigen_r);

    if (!strcmp(param->m_name, "df"))
    {
        // compute derivative of log probability wrt df:
        // lp_ddf=df*(dloggamma(df/2+1/2)-dloggamma(df/2))/2-1/2-
        // df*log(1+r2/(df*sigma^2))/2 +(df/2+1/2)*r2./(df*sigma^2+r2)
        eigen_r=(m_df*(CStatistics::dlgamma(m_df*0.5+0.5)-
            CStatistics::dlgamma(m_df*0.5))*0.5-0.5)*VectorXd::Ones(r.vlen);

        eigen_r-=m_df*(VectorXd::Ones(r.vlen)+
            eigen_r2/(m_df*CMath::sq(m_sigma))).array().log().matrix()/2.0;

        eigen_r+=(m_df/2.0+0.5)*eigen_r2.cwiseQuotient(
            eigen_r2+VectorXd::Ones(r.vlen)*(m_df*CMath::sq(m_sigma)));

        eigen_r*=(1.0-1.0/m_df);

        return r;
    }
    else if (!strcmp(param->m_name, "sigma"))
    {
        // compute derivative of log probability wrt sigma:
        // lp_dsigma=(df+1)*r2./a-1
        eigen_r=(m_df+1)*eigen_r2.cwiseQuotient(eigen_r2+
            VectorXd::Ones(r.vlen)*(m_df*CMath::sq(m_sigma)));
        eigen_r-=VectorXd::Ones(r.vlen);

        return r;
    }

    return SGVector<float64_t>();
}

SGVector<float64_t> CStudentsTLikelihood::get_second_derivative(const CLabels* lab,
        SGVector<float64_t> func, const TParameter* param) const
{
    // check the parameters
    REQUIRE(lab, "Labels are required (lab should not be NULL)\n")
    REQUIRE(lab->get_label_type()==LT_REGRESSION,
            "Labels must be type of CRegressionLabels\n")
    REQUIRE(lab->get_num_labels()==func.vlen, "Number of labels must match "
            "length of the function vector\n")

    SGVector<float64_t> r(func.vlen);
    Map<VectorXd> eigen_r(r.vector, r.vlen);

    Map<VectorXd> eigen_f(func.vector, func.vlen);

    SGVector<float64_t> y=((CRegressionLabels*)lab)->get_labels();
    Map<VectorXd> eigen_y(y.vector, y.vlen);

    // compute r=y-f and r2=(y-f).^2
    eigen_r=eigen_y-eigen_f;
    VectorXd eigen_r2=eigen_r.cwiseProduct(eigen_r);

    // compute a=r+sigma^2*df and a2=a.^2
    VectorXd a=eigen_r2+CMath::sq(m_sigma)*m_df*VectorXd::Ones(r.vlen);
    VectorXd a2=a.cwiseProduct(a);

    if (!strcmp(param->m_name, "df"))
    {
        // compute derivative of first derivative of log probability wrt df:
        // dlp_ddf=df*r.*(a-sigma^2*(df+1))./a2
        eigen_r=m_df*eigen_r.cwiseProduct(a-CMath::sq(m_sigma)*(m_df+1.0)*
            VectorXd::Ones(r.vlen)).cwiseQuotient(a2);
        eigen_r*=(1.0-1.0/m_df);

        return r;
    }
    else if (!strcmp(param->m_name, "sigma"))
    {
        // compute derivative of first derivative of log probability wrt sigma:
        // dlp_dsigma=-(df+1)*2*df*sigma^2*r./a2
        eigen_r=-(m_df+1.0)*2*m_df*CMath::sq(m_sigma)*
            eigen_r.cwiseQuotient(a2);

        return r;
    }

    return SGVector<float64_t>();
}

SGVector<float64_t> CStudentsTLikelihood::get_third_derivative(const CLabels* lab,
        SGVector<float64_t> func, const TParameter* param) const
{
    // check the parameters
    REQUIRE(lab, "Labels are required (lab should not be NULL)\n")
    REQUIRE(lab->get_label_type()==LT_REGRESSION,
            "Labels must be type of CRegressionLabels\n")
    REQUIRE(lab->get_num_labels()==func.vlen, "Number of labels must match "
            "length of the function vector\n")

    SGVector<float64_t> r(func.vlen);
    Map<VectorXd> eigen_r(r.vector, r.vlen);

    Map<VectorXd> eigen_f(func.vector, func.vlen);

    SGVector<float64_t> y=((CRegressionLabels*)lab)->get_labels();
    Map<VectorXd> eigen_y(y.vector, y.vlen);

    // compute r=y-f and r2=(y-f).^2
    eigen_r=eigen_y-eigen_f;
    VectorXd eigen_r2=eigen_r.cwiseProduct(eigen_r);

    // compute a=r+sigma^2*df and a3=a.^3
    VectorXd a=eigen_r2+CMath::sq(m_sigma)*m_df*VectorXd::Ones(r.vlen);
    VectorXd a3=(a.cwiseProduct(a)).cwiseProduct(a);

    if (!strcmp(param->m_name, "df"))
    {
        // compute derivative of second derivative of log probability wrt df:
        // d2lp_ddf=df*(r2.*(r2-3*sigma^2*(1+df))+df*sigma^4)./a3
        float64_t sigma2=CMath::sq(m_sigma);

        eigen_r=m_df*(eigen_r2.cwiseProduct(eigen_r2-3*sigma2*(1.0+m_df)*
            VectorXd::Ones(r.vlen))+(m_df*CMath::sq(sigma2))*VectorXd::Ones(r.vlen));
        eigen_r=eigen_r.cwiseQuotient(a3);

        eigen_r*=(1.0-1.0/m_df);

        return r;
    }
    else if (!strcmp(param->m_name, "sigma"))
    {
        // compute derivative of second derivative of log probability wrt sigma:
        // d2lp_dsigma=(df+1)*2*df*sigma^2*(a-4*r2)./a3
        eigen_r=(m_df+1.0)*2*m_df*CMath::sq(m_sigma)*
            (a-4.0*eigen_r2).cwiseQuotient(a3);

        return r;
    }

    return SGVector<float64_t>();
}

SGVector<float64_t> CStudentsTLikelihood::get_log_zeroth_moments(
        SGVector<float64_t> mu, SGVector<float64_t> s2, const CLabels* lab) const
{
    SGVector<float64_t> y;

    if (lab)
    {
        REQUIRE((mu.vlen==s2.vlen) && (mu.vlen==lab->get_num_labels()),
                "Length of the vector of means (%d), length of the vector of "
                "variances (%d) and number of labels (%d) should be the same\n",
                mu.vlen, s2.vlen, lab->get_num_labels())
        REQUIRE(lab->get_label_type()==LT_REGRESSION,
                "Labels must be type of CRegressionLabels\n")

        y=((CRegressionLabels*)lab)->get_labels();
    }
    else
    {
        REQUIRE(mu.vlen==s2.vlen, "Length of the vector of means (%d) and "
                "length of the vector of variances (%d) should be the same\n",
                mu.vlen, s2.vlen)

        y=SGVector<float64_t>(mu.vlen);
        y.set_const(1.0);
    }

    // create an object of normal pdf
    CNormalPDF* f=new CNormalPDF();

    // create an object of Student's t pdf
    CStudentsTPDF* g=new CStudentsTPDF();

    g->set_nu(m_df);
    g->set_sigma(m_sigma);

    // create an object of product of Student's-t pdf and normal pdf
    CProductFunction* h=new CProductFunction(f, g);
    SG_REF(h);

    // compute probabilities using numerical integration
    SGVector<float64_t> r(mu.vlen);

    for (index_t i=0; i<mu.vlen; i++)
    {
        // set normal pdf parameters
        f->set_mu(mu[i]);
        f->set_sigma(CMath::sqrt(s2[i]));

        // set Stundent's-t pdf parameters
        g->set_mu(y[i]);

        // evaluate integral on (-inf, inf)
        r[i]=CIntegration::integrate_quadgk(h, -CMath::INFTY, mu[i])+
            CIntegration::integrate_quadgk(h, mu[i], CMath::INFTY);
    }

    SG_UNREF(h);

    r.log();

    return r;
}

float64_t CStudentsTLikelihood::get_first_moment(SGVector<float64_t> mu,
        SGVector<float64_t> s2, const CLabels *lab, index_t i) const
{
    // check the parameters
    REQUIRE(lab, "Labels are required (lab should not be NULL)\n")
    REQUIRE((mu.vlen==s2.vlen) && (mu.vlen==lab->get_num_labels()),
            "Length of the vector of means (%d), length of the vector of "
            "variances (%d) and number of labels (%d) should be the same\n",
            mu.vlen, s2.vlen, lab->get_num_labels())
    REQUIRE(i>=0 && i<=mu.vlen, "Index (%d) out of bounds!\n", i)
    REQUIRE(lab->get_label_type()==LT_REGRESSION,
            "Labels must be type of CRegressionLabels\n")

    SGVector<float64_t> y=((CRegressionLabels*)lab)->get_labels();

    // create an object of normal pdf
    CNormalPDF* f=new CNormalPDF(mu[i], CMath::sqrt(s2[i]));

    // create an object of Student's t pdf
    CStudentsTPDF* g=new CStudentsTPDF(m_sigma, m_df, y[i]);

    // create an object of h(x)=N(x|mu,sigma)*t(x|mu,sigma,nu)
    CProductFunction* h=new CProductFunction(f, g);

    // create an object of k(x)=x*N(x|mu,sigma)*t(x|mu,sigma,nu)
    CProductFunction* k=new CProductFunction(new CLinearFunction(), h);
    SG_REF(k);

    // compute Z = \int N(x|mu,sigma)*t(x|mu,sigma,nu) dx
    float64_t Z=CIntegration::integrate_quadgk(h, -CMath::INFTY, mu[i])+
        CIntegration::integrate_quadgk(h, mu[i], CMath::INFTY);

    // compute 1st moment:
    // E[x] = Z^-1 * \int x*N(x|mu,sigma)*t(x|mu,sigma,nu)dx
    float64_t Ex=(CIntegration::integrate_quadgk(k, -CMath::INFTY, mu[i])+
            CIntegration::integrate_quadgk(k, mu[i], CMath::INFTY))/Z;

    SG_UNREF(k);

    return Ex;
}

float64_t CStudentsTLikelihood::get_second_moment(SGVector<float64_t> mu,
        SGVector<float64_t> s2, const CLabels *lab, index_t i) const
{
    // check the parameters
    REQUIRE(lab, "Labels are required (lab should not be NULL)\n")
    REQUIRE((mu.vlen==s2.vlen) && (mu.vlen==lab->get_num_labels()),
            "Length of the vector of means (%d), length of the vector of "
            "variances (%d) and number of labels (%d) should be the same\n",
            mu.vlen, s2.vlen, lab->get_num_labels())
    REQUIRE(i>=0 && i<=mu.vlen, "Index (%d) out of bounds!\n", i)
    REQUIRE(lab->get_label_type()==LT_REGRESSION,
            "Labels must be type of CRegressionLabels\n")

    SGVector<float64_t> y=((CRegressionLabels*)lab)->get_labels();

    // create an object of normal pdf
    CNormalPDF* f=new CNormalPDF(mu[i], CMath::sqrt(s2[i]));

    // create an object of Student's t pdf
    CStudentsTPDF* g=new CStudentsTPDF(m_sigma, m_df, y[i]);

    // create an object of h(x)=N(x|mu,sigma)*t(x|mu,sigma,nu)
    CProductFunction* h=new CProductFunction(f, g);

    // create an object of k(x)=x*N(x|mu,sigma)*t(x|mu,sigma,nu)
    CProductFunction* k=new CProductFunction(new CLinearFunction(), h);
    SG_REF(k);

    // create an object of p(x)=x^2*N(x|mu,sigma^2)*t(x|mu,sigma,nu)
    CProductFunction* p=new CProductFunction(new CQuadraticFunction(), h);
    SG_REF(p);

    // compute Z = \int N(x|mu,sigma)*t(x|mu,sigma,nu) dx
    float64_t Z=CIntegration::integrate_quadgk(h, -CMath::INFTY, mu[i])+
        CIntegration::integrate_quadgk(h, mu[i], CMath::INFTY);

    // compute 1st moment:
    // E[x] = Z^-1 * \int x*N(x|mu,sigma)*t(x|mu,sigma,nu)dx
    float64_t Ex=(CIntegration::integrate_quadgk(k, -CMath::INFTY, mu[i])+
            CIntegration::integrate_quadgk(k, mu[i], CMath::INFTY))/Z;

    // compute E[x^2] = Z^-1 * \int x^2*N(x|mu,sigma)*t(x|mu,sigma,nu)dx
    float64_t Ex2=(CIntegration::integrate_quadgk(p, -CMath::INFTY, mu[i])+
            CIntegration::integrate_quadgk(p, mu[i], CMath::INFTY))/Z;

    SG_UNREF(k);
    SG_UNREF(p);

    // return 2nd moment: Var[x]=E[x^2]-E[x]^2
    return Ex2-CMath::sq(Ex);
}
}

#endif /* HAVE_EIGEN3 */