statlib.c

使用具有增量学习的监控式学习方法。包括几个不同的分类算法。

/**
 * @file
 * Test statistics.
 * Calculates statistics and print plot data.
 *
 * @author  Mikael Ylikoski
 * @date    2001-2002
 */

#include <stdio.h>
#include <string.h>
#include "statlib.h"
#include "utility.h"

/**
 * Box statistics.
 */
typedef struct {
    int notm;		/**< Number Of Trained Messages */
    int nopm;		/**< Number Of classified messages */
    int *stats;		/**< Classification statistics */
    int tp;		/**< a = messages correctly assigned to it */
    int fp;		/**< b = messages incorrectly assigned to it */
    int fn;		/**< c = messages incorrectly rejected from it */
    int tn;		/**< d = messages correctly rejected from it */
} box_stat;

/**
 * Classifier statistics.
 */
typedef struct {
    int nocm;		/**< Number Of Correctly classified Messages */
    int nolm;		/**< Number Of correctly Limited Messages */
    int nofm;		/**< Number Of Falsely classified Messages */
} cls_stat;

static enum plotting_mode plot_mode;	/**< Plotting mode */

static cls_stat *cls_stats;		/**< Classifier statistics */
static int noc;				/**< Number Of Classifiers */

//static box_stat *box_stats;		/**< Box statistics */
static int nob;				/**< Number Of Boxes */

static int plot_c;			/**< Plot classifier */
static int notm;			/**< Number Of Trained Messages */

static int *n_list;			/**< N-average list */
static int mean_n;			/**< Number of steps in average */
static int non;				/**< Number Of N-average messages */
static int nocn;			/**< Number Of Correctly classified
					     N-average messages */

static int rank_limit;			/**< Threshold for accepted rank */


/**
 * Update N-average statistics.
 *
 * @param b  1 for correct classification, 0 otherwise
 * @return Zero if ok, non-zero otherwise.
 */
static int
update_n_stat (int b) {
    int i;

    if (plot_mode != N_AVERAGE)
	return 0;

    if (notm <= mean_n)
	non = notm;
    else
	non = mean_n;

    if (b)
	b = 1;

    i = (notm - 1) % mean_n;
    if (n_list[i] != b) {
	n_list[i] = b;
	if (b)
	    nocn++;
	else
	    nocn--;
    }

    return 0;
}


/**
 * Update statistics for new test message.
 *
 * @param td  test data
 * @return Zero if ok, or nonzero otherwise.
 */
int
statlib_update_stats (test_data *td) {
    int i, k, ok;

    if (td->bno == -1 || td->mno == -1)
	return -1;

    if (td->nor > noc) {
	//cls_stats = my_realloc (cls_stats, td->nor * sizeof(cls_stat));
	fprintf (stderr, "Error: Wrong number of classifiers.\n");
	return -1;
    }

    notm++;
    ok = 0;
    for (k = 0; k < td->nor && k < noc; k++) {
	switch (td->res[k].cm) {
	case RANK:
	    if (td->res[k].len < 1)
		break;

	    for (i = 0; i < rank_limit && i < td->res[k].len; i++)
		if (td->res[k].u.rlist[i] == td->bno) {
		    cls_stats[k].nolm++;
		    if (k == plot_c)
			ok = 1;
		    break;
		}

	    if (td->res[k].u.rlist[0] == td->bno)
		cls_stats[k].nocm++;
	    else
		cls_stats[k].nofm++;
	    break;
	case SCORE:
	    return -1;
	}
    }

    update_n_stat (ok);

    return 0;
}


/**
 * Print header.
 */
void
statlib_print_plot_header (void) {
    printf ("# Process this data with gnuplot to create a plot.\n"
	    "# gnuplot> plot \"datafile\" notitle with lines\n");
}


/**
 * Print current data for plot.
 */
void
statlib_print_plot_data (void) {
    printf ("%d ", notm);
    switch (plot_mode) {
    case TOTAL:
	printf ("%.3f\n", cls_stats[plot_c].nolm / (double)notm);
	break;
    case N_AVERAGE:
	printf ("%.3f\n", nocn / (double)non);
	break;
    }
}


/**
 * Print statistics.
 */
void
statlib_print_results (void) {
    int i, j;

    for (i = 0; i < noc; i++) {
	printf ("## Classifier %d", i);
	if (i == plot_c)
	    printf (" (plotted)");
	printf ("\n");
	j = cls_stats[i].nocm + cls_stats[i].nofm;
	printf ("##   Correct: %d  Top-%d-Correct: %d  Covered: %d  "
		"Total: %d\n", cls_stats[i].nocm, rank_limit,
		cls_stats[i].nolm, j, notm);
	if (j)
	    printf ("##   Accuracy: %.3f  Top-%d-Accuracy: %.3f  "
		    "Coverage: %.3f\n",
		    cls_stats[i].nocm / (double)j, rank_limit,
		    cls_stats[i].nolm / (double)j, j / (double)notm);
	else
	    printf ("##   Accuracy: 0      Top-%d-Accuracy: 0      "
		    "Coverage: 0\n", rank_limit);
	printf ("##   Total-Accuracy: %.3f  Total-Top-%d-Accuracy: %.3f\n",
		cls_stats[i].nocm / (double)notm, rank_limit,
		cls_stats[i].nolm / (double)notm);
    }
}


/**
 * Initialization.
 *
 * @return Zero if ok, or nonzero otherwise.
 */
int
statlib_initialize (enum plotting_mode in_plm, int in_rklm, int in_mean,
		    int in_plotc, int in_noc, int in_nob) {
    int i;

    /* Configuration */
    plot_mode = in_plm;
    rank_limit = in_rklm;
    mean_n = in_mean;
    plot_c = in_plotc;
    noc = in_noc;
    nob = in_nob;

    /* Allocate statistics memory */
    cls_stats = my_malloc (noc * sizeof(cls_stat));
    for (i = 0; i < noc; i++) {
	cls_stats[i].nocm = 0;
	cls_stats[i].nolm = 0;
	cls_stats[i].nofm = 0;
    }

    if (plot_mode == N_AVERAGE)
	n_list = my_calloc (mean_n, sizeof(int));

    non = 0;
    nocn = 0;
    notm = 0;

    return 0;
}