ROCEvaluation.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/evaluation/ROCEvaluation.h>
#include <shogun/mathematics/Math.h>

using namespace shogun;

CROCEvaluation::~CROCEvaluation()
{
}

float64_t CROCEvaluation::evaluate(CLabels* predicted, CLabels* ground_truth)
{
    return evaluate_roc(predicted,ground_truth);
}

float64_t CROCEvaluation::evaluate_roc(CLabels* predicted, CLabels* ground_truth)
{
    ASSERT(predicted && ground_truth);
    ASSERT(predicted->get_num_labels()==ground_truth->get_num_labels());
    ASSERT(predicted->get_label_type()==LT_BINARY);
    ASSERT(ground_truth->get_label_type()==LT_BINARY);
    ground_truth->ensure_valid();

    // assume threshold as negative infinity
    float64_t threshold = CMath::ALMOST_NEG_INFTY;
    // false positive rate
    float64_t fp = 0.0;
    // true positive rate
    float64_t tp=0.0;

    int32_t i;
    // total number of positive labels in predicted
    int32_t pos_count=0;
    int32_t neg_count=0;

    // initialize number of labels and labels
    SGVector<float64_t> orig_labels(predicted->get_num_labels());
    int32_t length = orig_labels.vlen;
    for (i=0; i<length; i++)
        orig_labels[i] = predicted->get_confidence(i);
    float64_t* labels = SGVector<float64_t>::clone_vector(orig_labels.vector, length);

    // get sorted indexes
    int32_t* idxs = SG_MALLOC(int32_t, length);
    for(i=0; i<length; i++)
        idxs[i] = i;

    CMath::qsort_backward_index(labels,idxs,length);

    // number of different predicted labels
    int32_t diff_count=1;

    // get number of different labels
    for (i=0; i<length-1; i++)
    {
        if (labels[i] != labels[i+1])
            diff_count++;
    }

    SG_FREE(labels);

    // initialize graph and auROC
    m_ROC_graph = SGMatrix<float64_t>(2,diff_count+1);
    m_thresholds = SGVector<float64_t>(length);
    m_auROC = 0.0;

    // get total numbers of positive and negative labels
    for(i=0; i<length; i++)
    {
        if (ground_truth->get_confidence(i) >= 0)
            pos_count++;
        else
            neg_count++;
    }

    // assure both number of positive and negative examples is >0
    ASSERT(pos_count>0 && neg_count>0);

    int32_t j = 0;
    float64_t label;

    // create ROC curve and calculate auROC
    for(i=0; i<length; i++)
    {
        label = predicted->get_confidence(idxs[i]);

        if (label != threshold)
        {
            threshold = label;
            m_ROC_graph[2*j] = fp/neg_count;
            m_ROC_graph[2*j+1] = tp/pos_count;
            j++;
        }

        m_thresholds[i]=threshold;

        if (ground_truth->get_confidence(idxs[i]) > 0)
            tp+=1.0;
        else
            fp+=1.0;
    }

    // add (1,1) to ROC curve
    m_ROC_graph[2*diff_count] = 1.0;
    m_ROC_graph[2*diff_count+1] = 1.0;

    // calc auROC using area under curve
    m_auROC = CMath::area_under_curve(m_ROC_graph.matrix,diff_count+1,false);

    m_computed = true;

    return m_auROC;
}

SGMatrix<float64_t> CROCEvaluation::get_ROC()
{
    if (!m_computed)
        SG_ERROR("Uninitialized, please call evaluate first");

    return m_ROC_graph;
}

SGVector<float64_t> CROCEvaluation::get_thresholds()
{
    if (!m_computed)
        SG_ERROR("Uninitialized, please call evaluate first");

    return m_thresholds;
}

float64_t CROCEvaluation::get_auROC()
{
    if (!m_computed)
            SG_ERROR("Uninitialized, please call evaluate first");

    return m_auROC;
}