MultitaskKernelPlifNormalizer.h

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 2 of the License, or
 * (at your option) any later version.
 *
 * Written (W) 2010 Christian Widmer
 * Copyright (C) 2010 Max-Planck-Society
 */

#ifndef _MULTITASKKERNELPLIFNORMALIZER_H___
#define _MULTITASKKERNELPLIFNORMALIZER_H___

#include "kernel/KernelNormalizer.h"
#include "kernel/MultitaskKernelMklNormalizer.h"
#include "kernel/Kernel.h"
#include <algorithm>



namespace shogun
{
class CMultitaskKernelPlifNormalizer: public CMultitaskKernelMklNormalizer
{

public:

    CMultitaskKernelPlifNormalizer(std::vector<float64_t> support_, std::vector<int32_t> task_vector)
    {

        num_betas = static_cast<int>(support_.size());

        support = support_;

        // init support points values with constant function
        betas = std::vector<float64_t>(num_betas);
        for (int i=0; i!=num_betas; i++)
        {
            betas[i] = 1;
        }

        num_tasks = get_num_unique_tasks(task_vector);

        // set both sides equally
        set_task_vector(task_vector);

        // init distance matrix
        distance_matrix = std::vector<float64_t>(num_tasks * num_tasks);

        // init similarity matrix
        similarity_matrix = std::vector<float64_t>(num_tasks * num_tasks);

    }


    inline virtual float64_t normalize(float64_t value, int32_t idx_lhs,
            int32_t idx_rhs)
    {

        //lookup tasks
        int32_t task_idx_lhs = task_vector_lhs[idx_lhs];
        int32_t task_idx_rhs = task_vector_rhs[idx_rhs];

        //lookup similarity
        float64_t task_similarity = get_task_similarity(task_idx_lhs,
                task_idx_rhs);

        //take task similarity into account
        float64_t similarity = (value/scale) * task_similarity;


        return similarity;

    }

    int32_t get_num_unique_tasks(std::vector<int32_t> vec) {

        //sort
        std::sort(vec.begin(), vec.end());

        //reorder tasks with unique prefix
        std::vector<int32_t>::iterator endLocation = std::unique(vec.begin(), vec.end());

        //count unique tasks
        int32_t num_vec = std::distance(vec.begin(), endLocation);

        return num_vec;

    }


    virtual ~CMultitaskKernelPlifNormalizer()
    {
    }


    void update_cache()
    {


        for (int32_t i=0; i!=num_tasks; i++)
        {
            for (int32_t j=0; j!=num_tasks; j++)
            {

                float64_t similarity = compute_task_similarity(i, j);
                set_task_similarity(i,j,similarity);

            }

        }
    }


    float64_t compute_task_similarity(int32_t task_a, int32_t task_b)
    {

        float64_t distance = get_task_distance(task_a, task_b);
        float64_t similarity = -1;

        int32_t upper_bound_idx = -1;


        // determine interval
        for (int i=1; i!=num_betas; i++)
        {
            if (distance <= support[i])
            {
                upper_bound_idx = i;
                break;
            }
        }

        // perform interpolation (constant for beyond upper bound)
        if (upper_bound_idx == -1)
        {

            similarity = betas[num_betas-1];

        } else {

            int32_t lower_bound_idx = upper_bound_idx - 1;
            float64_t interval_size = support[upper_bound_idx] - support[lower_bound_idx];

            float64_t factor_lower = 1 - (distance - support[lower_bound_idx]) / interval_size;
            float64_t factor_upper = 1 - factor_lower;

            similarity = factor_lower*betas[lower_bound_idx] + factor_upper*betas[upper_bound_idx];

        }

        return similarity;

    }


public:

    virtual std::vector<int32_t> get_task_vector_lhs() const
    {
        return task_vector_lhs;
    }

    virtual void set_task_vector_lhs(std::vector<int32_t> vec)
    {
        task_vector_lhs = vec;
    }

    virtual std::vector<int32_t> get_task_vector_rhs() const
    {
        return task_vector_rhs;
    }

    virtual void set_task_vector_rhs(std::vector<int32_t> vec)
    {
        task_vector_rhs = vec;
    }

    virtual void set_task_vector(std::vector<int32_t> vec)
    {
        task_vector_lhs = vec;
        task_vector_rhs = vec;
    }

    float64_t get_task_distance(int32_t task_lhs, int32_t task_rhs)
    {

        ASSERT(task_lhs < num_tasks && task_lhs >= 0);
        ASSERT(task_rhs < num_tasks && task_rhs >= 0);

        return distance_matrix[task_lhs * num_tasks + task_rhs];

    }

    void set_task_distance(int32_t task_lhs, int32_t task_rhs,
            float64_t distance)
    {

        ASSERT(task_lhs < num_tasks && task_lhs >= 0);
        ASSERT(task_rhs < num_tasks && task_rhs >= 0);

        distance_matrix[task_lhs * num_tasks + task_rhs] = distance;

    }

    float64_t get_task_similarity(int32_t task_lhs, int32_t task_rhs)
    {

        ASSERT(task_lhs < num_tasks && task_lhs >= 0);
        ASSERT(task_rhs < num_tasks && task_rhs >= 0);

        return similarity_matrix[task_lhs * num_tasks + task_rhs];

    }

    void set_task_similarity(int32_t task_lhs, int32_t task_rhs,
            float64_t similarity)
    {

        ASSERT(task_lhs < num_tasks && task_lhs >= 0);
        ASSERT(task_rhs < num_tasks && task_rhs >= 0);

        similarity_matrix[task_lhs * num_tasks + task_rhs] = similarity;

    }

    float64_t get_beta(int32_t idx)
    {

        return betas[idx];

    }

    void set_beta(int32_t idx, float64_t weight)
    {

        betas[idx] = weight;

        update_cache();

    }

    int32_t get_num_betas()
    {

        return num_betas;

    }


    inline virtual const char* get_name() const
    {
        return "MultitaskKernelNormalizer";
    }

protected:

    int32_t num_tasks;

    std::vector<int32_t> task_vector_lhs;

    std::vector<int32_t> task_vector_rhs;

    std::vector<float64_t> distance_matrix;

    std::vector<float64_t> similarity_matrix;

    int32_t num_betas;

    std::vector<float64_t> betas;

    std::vector<float64_t> support;

};
}
#endif