/* compute mean, maximum, minimum, difference between min & max, variance, median,
1st quartile, 3rd quartile, interquartile range of a list of real values */
int getStatistics(mx_real_t* features, mx_real_t* list,int count) {
int i;
mx_real_t *_list;
mx_real_t mean=0, max=-MX_REAL_MAX, min=MX_REAL_MAX, variance=0;
mx_real_t median, quartile_1, quartile_3;
if (count==0) {
for (i=0;i<STATS;i++)
features[i]=0.0;
return STATS;
}
if (count==1) {
features[0]=features[1]=features[2]=features[5]=features[6]=features[7]=list[0];
features[3]=features[4]=features[8]=0.0;
return STATS;
}
_list= (mx_real_t *) rs_malloc(count*sizeof(mx_real_t),"statistics list");
for (i=0;i<count;i++) {
_list[i]=list[i];
mean+=_list[i];
variance+=mx_sqr(_list[i]);
if (_list[i] <min)
min=_list[i];
if (_list[i] >max)
max=_list[i];
}
mean /= count;
variance = variance/count - mx_sqr(mean);
qsort(_list,count,sizeof(mx_real_t),_cmp_mx_real);
median = _list[count/2];
quartile_1 = _list[count/4];
quartile_3 = _list[3*count/4];
features[0] = mean;
features[1] = max;
features[2] = min;
features[3] = max - min;
features[4] = variance;
features[5] = median;
features[6] = quartile_1;
features[7] = quartile_3;
features[8] = fabs(quartile_3 - quartile_1);
rs_free(_list);
return STATS;
}
naive_bayes_classifier_t *nB_update_classifier(naive_bayes_classifier_t *nB, int class_ind, mx_real_t *features) {
int i;
mx_real_t temp;
nB->n_instances[class_ind]++;
if (!nB->finished) {
for (i=0;i<nB->feature_dim;i++) {
nB->means[class_ind][i]+=features[i];
nB->std_dev[class_ind][i]+=features[i]*features[i];
}
}
else {
for (i=0;i<nB->feature_dim;i++) {
temp = nB->means[class_ind][i];
nB->means[class_ind][i]=(nB->means[class_ind][i]*nB->n_instances[class_ind] + features[i]) / (nB->n_instances[class_ind]+ 1.0);
nB->std_dev[class_ind][i] = sqrt(((mx_sqr(nB->std_dev[class_ind][i]) + temp * temp)*nB->n_instances[class_ind] + features[i]*features[i]) / (nB->n_instances[class_ind]+1.0) - nB->means[class_ind][i]*nB->means[class_ind][i]);
}
for (i=0;i<nB->n_classes; i++) {
nB->class_probs[i]= 1.0*nB->n_instances[i]/(nB->n_instances[nB->n_classes]+1);
}
}
nB->n_instances[nB->n_classes]++;
return nB;
}
/* Standard normal distribution; often results in 0 for many attributes */
mx_real_t nB_class_prob_simple(naive_bayes_classifier_t *nB, mx_real_t *instance, int class_ind) {
int j;
mx_real_t prob=0;
mx_real_t diff, temp, stddev;
mx_real_t norm_const=sqrt(2 * M_PI);
prob = nB->class_probs[class_ind];
for (j=0; j<nB->feature_dim; j++) {
diff = instance[j]-nB->means[class_ind][j];
stddev=nB->std_dev[class_ind][j];
if (stddev ==0)
stddev=MX_REAL_MIN;
temp = (1 / (norm_const * stddev)) * exp(-(mx_sqr(diff) / (2 * mx_sqr(stddev))));
prob *= temp;
}
return prob;
}
/* log normal distribution; can cope with many attribute, but difficult to normalise */
mx_real_t nB_class_prob(naive_bayes_classifier_t *nB, mx_real_t *instance, int class_ind) {
int j;
mx_real_t prob=0, sqr;
prob = mx_log(nB->class_probs[class_ind]);
for (j=0; j<nB->feature_dim; j++) {
// stdev == 0 -> attribute is constant
if (nB->std_dev[class_ind][j]==0)
continue;
sqr = mx_sqr(nB->std_dev[class_ind][j]);
if (sqr !=0)
prob+=-mx_log(nB->std_dev[class_ind][j])-mx_sqr(instance[j]-nB->means[class_ind][j])/(2*sqr);
else
prob+=-mx_log(nB->std_dev[class_ind][j])-mx_sqr(instance[j]-nB->means[class_ind][j])/(2*MX_REAL_MIN);
}
// Transformation in vernünftigen Wertebereich
return exp(prob/nB->feature_dim);
}