selector.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。

/* Copyright (C) 2002  Mikael Ylikoski
 * See the accompanying file "README" for the full copyright notice */
/**
 * @file
 * Feature selector. (s/selector/weighter/ ?)
 *
 * normalizer: cannot learn, can change weights on vectors
 * weighter: can learn, can change weights on vectors
 * selector: can learn, can remove features from vectorizer dictionary
 *
 * @author  Mikael Ylikoski
 * @date    2002
 */

#include <math.h>
#include <stdlib.h>
#include <string.h>
#include "utility.h"
#include "vector.h"
#include "selector.h"


/**
 * Feature selector.
 */
struct selector_ {
    //vector *tdf;	/**< Term Document Frequency vector */
    //int *tcf;		/**< Term Class Frequency */
    int nod;		/**< Number Of Documents */
    //int noc;		/**< Number Of Classes */

    int *tf;		/**< Term Frequency */
    int *tb;		/**< Term Birthtime */
    int not;		/**< Number Of Terms */
};


/**
 * Create a new selector.
 *
 * @return A new selector.
 */
selector *
selector_new (void) {
    selector *sel;

    sel = my_malloc (sizeof(selector));
    //sel->tdf = vector_new (100);
    //sel->tcf = my_calloc (1, sizeof(int));
    //sel->noc = 1;
    sel->nod = 0;

    sel->tf = NULL;
    sel->tb = NULL;
    sel->not = 0;

    return sel;
}


/**
 * Update the selector statistics.
 *
 * @param sel  selector to update
 */
int
selector_update (selector *sel, vector *v, int class) {
    int i, j, k;

    //vector_add_v (sel->tdf, v, 1);
    //sel->tcf[class]++;
    sel->nod++;

    if (v->nel > 0) {
	j = v->name[v->nel - 1];
	if (j >= sel->not) {
	    k = j + 1;
	    sel->tf = my_realloc (sel->tf, k * sizeof(int));
	    sel->tb = my_realloc (sel->tb, k * sizeof(int));
	    memset (&sel->tf[sel->not], 0, (k - sel->not) * sizeof(int));
	    memset (&sel->tb[sel->not], 0, (k - sel->not) * sizeof(int));
	    sel->not = k;
	}
    }

    for (i = 0; i < v->nel; i++) {
	j = v->name[i];
	if (sel->tf[j] == 0)
	    sel->tb[j] = sel->nod;
	sel->tf[j]++;	// += v->value[i]
    }

    return 0;
}


/**
 * Reweight a vector.
 */
void
selector_weight (selector *sel, vector *v) {
}


/**
 * Used by vectorizer to remove terms.
 *
 * @param sel  selector to use
 * @return A boolean vector with ones for all features to remove
 *         (if autobias == 1 then feature 0 may never be removed)
 */
vector *
selector_get_removable (selector *sel) {
    int i;
    vector *v;

    v = vector_new (10);
    for (i = 1; i < sel->not; i++)
	if (sel->tf[i] && sel->tf[i] <
	    log (sel->nod - sel->tb[i]) * 1 - 1) {
			   /* ifile use * (1 / log (2)) = 1.4427 */
	    sel->tf[i] = 0;
	    vector_append (v, i);
	}
    if (v->nel < 1) {
	vector_free (v);
	v = NULL;
    } //else
    //vector_print (v);

    return v;
}