NormalizeKernelUnitVariance.h

Go to the documentation of this file.

//===========================================================================
/*!
 * 
 *
 * \brief       Determine the scaling factor of a ScaledKernel so that it has unit variance in feature space one on a given dataset.
 * 
 * 
 * 
 *
 * \author      M. Tuma
 * \date        2012
 *
 *
 * \par Copyright 1995-2017 Shark Development Team
 * 
 * <BR><HR>
 * This file is part of Shark.
 * <https://shark-ml.github.io/Shark/>
 * 
 * Shark is free software: you can redistribute it and/or modify
 * it under the terms of 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.
 * 
 * Shark is distributed in the hope that it will be useful,
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
 * GNU Lesser General Public License for more details.
 * 
 * You should have received a copy of the GNU Lesser General Public License
 * along with Shark.  If not, see <http://www.gnu.org/licenses/>.
 *
 */
//===========================================================================
 
 
#ifndef SHARK_ALGORITHMS_TRAINERS_NORMALIZEKERNELUNITVARIANCE_H
#define SHARK_ALGORITHMS_TRAINERS_NORMALIZEKERNELUNITVARIANCE_H
 
 
#include <shark/Models/Kernels/ScaledKernel.h>
#include <shark/Algorithms/Trainers/AbstractTrainer.h>
 
namespace shark {
 
 
///
/// \brief Determine the scaling factor of a ScaledKernel so that it has unit variance in feature space one on a given dataset.
///
/// \par
/// For example in the multiple kernel learning setting, it can be important that the sub-kernels
/// are normalized to unit variance in feature space. This class computes both the trace and the
/// mean of a kernel matrix, and uses both to employ the "Multiplicative Kernel Scaling" laid out
/// in "Kloft, Brefeld, Sonnenburg, Zien: l_p-Norm Multiple Kernel Learning. JMLR 12, 2011.".
/// Given a ScaledKernel, which itself holds an arbitrary underlying kernel k, we compute
/// \f[ \frac{1}{N}\sum_{i=1}^N k(x_i,x_i) - \frac{1}{N^2} \sum_{i,j=1}^N k(x_i,x_j) \f]
/// 
/// \ingroup unsupervised_trainer
template < class InputType = RealVector >
class NormalizeKernelUnitVariance : public AbstractUnsupervisedTrainer<ScaledKernel<InputType> >
{
public:
 
    NormalizeKernelUnitVariance()
    { }
 
    /// \brief From INameable: return the class name.
    std::string name() const
    { return "NormalizeKernelUnitVariance"; }
 
    double trace() const {
        return m_matrixTrace;
    }
    double mean() const {
        return m_mean;
    }
 
    void train( ScaledKernel<InputType>& kernel, UnlabeledData<InputType> const& input )
    {
        SHARK_RUNTIME_CHECK(input.numberOfElements() >= 2, "Input needs to contain at least two points");
        AbstractKernelFunction< InputType > const& k = *kernel.base(); //get direct access to the kernel we want to use.        
        
        // Next compute the trace and mean of the kernel matrix. This means heavy lifting: 
        // calculate each entry of one diagonal half of the kernel matrix exactly once.
        // [MT] This part was taken from the PrecomputedMatrix constructor in QuadraticProgram.h
        // Blockwise version, with order of computations optimized for better use of the processor
        // cache. This can save around 10% computation time for fast kernel functions.
        std::size_t N = input.numberOfElements();
        
        //O.K. tried to make it more efficient (and shorter)
        m_mean = 0.0;
        m_matrixTrace = 0.0;
        for(std::size_t i = 0; i != input.numberOfBatches(); ++i){
            typename UnlabeledData<InputType>::const_batch_reference batch = input.batch(i);
            //off diagonal entries
            for(std::size_t j = 0; j < i; ++j){
                RealMatrix matrixBlock = k(batch, input.batch(j));
                m_mean += 2*sum(matrixBlock);
            }
            RealMatrix matrixBlock = k(batch, batch);
            m_mean += sum(matrixBlock);
            m_matrixTrace += blas::trace(matrixBlock);
        }
        double tm = m_matrixTrace/N - m_mean/N/N;
        SHARK_ASSERT( tm > 0 );
        double scaling_factor = 1.0 / tm;
        kernel.setFactor( scaling_factor );
    }
 
protected:
    double m_mean; //store for other uses, external queries, etc.
    double m_matrixTrace;
};
 
 
}
#endif