autoClassificationTable Discriminant_and_TableOfReal_to_ClassificationTable_dw (Discriminant me, TableOfReal thee, int poolCovarianceMatrices, int useAprioriProbabilities, double alpha, double minProb, autoTableOfReal *displacements) {
try {
long g = Discriminant_getNumberOfGroups (me);
long p = Eigen_getDimensionOfComponents (my eigen.get());
long m = thy numberOfRows;
if (p != thy numberOfColumns) Melder_throw
(U"The number of columns does not agree with the dimension of the discriminant.");
autoNUMvector<double> log_p (1, g);
autoNUMvector<double> log_apriori (1, g);
autoNUMvector<double> ln_determinant (1, g);
autoNUMvector<double> buf (1, p);
autoNUMvector<double> displacement (1, p);
autoNUMvector<double> x (1, p);
autoNUMvector<SSCP> sscpvec (1, g);
autoSSCP pool = SSCPList_to_SSCP_pool (my groups.get());
autoClassificationTable him = ClassificationTable_create (m, g);
NUMstrings_copyElements (thy rowLabels, his rowLabels, 1, m);
autoTableOfReal adisplacements = Data_copy (thee);
// Scale the sscp to become a covariance matrix.
for (long i = 1; i <= p; i ++) {
for (long k = i; k <= p; k ++) {
pool -> data [k] [i] = pool -> data [i] [k] /= pool -> numberOfObservations - g;
}
}
double lnd;
autoSSCPList agroups;
SSCPList groups;
if (poolCovarianceMatrices) {
// Covariance matrix S can be decomposed as S = L.L'. Calculate L^-1.
// L^-1 will be used later in the Mahalanobis distance calculation:
// v'.S^-1.v == v'.L^-1'.L^-1.v == (L^-1.v)'.(L^-1.v).
NUMlowerCholeskyInverse (pool -> data, p, & lnd);
for (long j = 1; j <= g; j ++) {
ln_determinant [j] = lnd;
sscpvec [j] = pool.get();
}
groups = my groups.get();
} else {
//Calculate the inverses of all group covariance matrices.
// In case of a singular matrix, substitute inverse of pooled.
agroups = Data_copy (my groups.get());
groups = agroups.get();
long npool = 0;
for (long j = 1; j <= g; j ++) {
SSCP t = groups->at [j];
long no = (long) floor (SSCP_getNumberOfObservations (t));
for (long i = 1; i <= p; i ++) {
for (long k = i; k <= p; k ++) {
t -> data [k] [i] = t -> data [i] [k] /= no - 1;
}
}
sscpvec [j] = groups->at [j];
try {
NUMlowerCholeskyInverse (t -> data, p, & ln_determinant [j]);
} catch (MelderError) {
// Try the alternative: the pooled covariance matrix.
// Clear the error.
Melder_clearError ();
if (npool == 0) {
NUMlowerCholeskyInverse (pool -> data, p, & lnd);
}
npool ++;
sscpvec [j] = pool.get();
ln_determinant [j] = lnd;
}
}
if (npool > 0) {
Melder_warning (npool, U" groups use pooled covariance matrix.");
}
}
// Labels for columns in ClassificationTable
for (long j = 1; j <= g; j ++) {
const char32 *name = Thing_getName (my groups->at [j]);
if (! name) {
name = U"?";
}
TableOfReal_setColumnLabel (him.get(), j, name);
}
// Normalize the sum of the apriori probabilities to 1.
// Next take ln (p) because otherwise probabilities might be too small to represent.
double logg = log (g);
NUMvector_normalize1 (my aprioriProbabilities, g);
for (long j = 1; j <= g; j ++) {
log_apriori[j] = ( useAprioriProbabilities ? log (my aprioriProbabilities[j]) : - logg );
}
// Generalized squared distance function:
// D^2(x) = (x - mu)' S^-1 (x - mu) + ln (determinant(S)) - 2 ln (apriori)
for (long i = 1; i <= m; i ++) {
SSCP winner;
double norm = 0, pt_max = -1e308;
long iwinner = 1;
for (long k = 1; k <= p; k ++) {
x [k] = thy data [i] [k] + displacement [k];
}
for (long j = 1; j <= g; j ++) {
SSCP t = groups->at [j];
double md = mahalanobisDistanceSq (sscpvec [j] -> data, p, x.peek(), t -> centroid, buf.peek());
double pt = log_apriori [j] - 0.5 * (ln_determinant [j] + md);
if (pt > pt_max) {
pt_max = pt;
iwinner = j;
}
log_p [j] = pt;
}
for (long j = 1; j <= g; j ++) {
norm += log_p [j] = exp (log_p [j] - pt_max);
}
for (long j = 1; j <= g; j ++) {
his data [i] [j] = log_p [j] / norm;
}
// Save old displacement, calculate new displacement
winner = groups->at [iwinner];
for (long k = 1; k <= p; k ++) {
adisplacements -> data [i] [k] = displacement [k];
if (his data [i] [iwinner] > minProb) {
double delta_k = winner -> centroid [k] - x [k];
displacement [k] += alpha * delta_k;
}
}
}
*displacements = adisplacements.move();
return him;
} catch (MelderError) {
Melder_throw (U"ClassificationTable for Weenink procedure not created.");
}
}
ClassificationTable Discriminant_and_TableOfReal_to_ClassificationTable (Discriminant me, TableOfReal thee,
int poolCovarianceMatrices, int useAprioriProbabilities) {
try {
long g = Discriminant_getNumberOfGroups (me);
long p = Eigen_getDimensionOfComponents (me);
long m = thy numberOfRows;
if (p != thy numberOfColumns) Melder_throw
(U"The number of columns does not agree with the dimension of the discriminant.");
autoNUMvector<double> log_p (1, g);
autoNUMvector<double> log_apriori (1, g);
autoNUMvector<double> ln_determinant (1, g);
autoNUMvector<double> buf (1, p);
autoNUMvector<SSCP> sscpvec (1, g);
autoSSCP pool = SSCPs_to_SSCP_pool (my groups);
autoClassificationTable him = ClassificationTable_create (m, g);
NUMstrings_copyElements (thy rowLabels, his rowLabels, 1, m);
// Scale the sscp to become a covariance matrix.
for (long i = 1; i <= p; i++) {
for (long k = i; k <= p; k++) {
pool -> data[k][i] = (pool -> data[i][k] /= (pool -> numberOfObservations - g));
}
}
double lnd;
autoSSCPs agroups = 0; SSCPs groups;
if (poolCovarianceMatrices) {
/*
Covariance matrix S can be decomposed as S = L.L'. Calculate L^-1.
L^-1 will be used later in the Mahalanobis distance calculation:
v'.S^-1.v == v'.L^-1'.L^-1.v == (L^-1.v)'.(L^-1.v).
*/
NUMlowerCholeskyInverse (pool -> data, p, &lnd);
for (long j = 1; j <= g; j++) {
ln_determinant[j] = lnd;
sscpvec[j] = pool.peek();
}
groups = (SSCPs) my groups;
} else {
// Calculate the inverses of all group covariance matrices.
// In case of a singular matrix, substitute inverse of pooled.
agroups.reset (Data_copy ( (SSCPs) my groups)); groups = agroups.peek();
long npool = 0;
for (long j = 1; j <= g; j++) {
SSCP t = (SSCP) groups -> item[j];
long no = (long) floor (SSCP_getNumberOfObservations (t));
for (long i = 1; i <= p; i++) {
for (long k = i; k <= p; k++) {
t -> data[k][i] = (t -> data[i][k] /= (no - 1));
}
}
sscpvec[j] = (SSCP) groups -> item[j];
try {
NUMlowerCholeskyInverse (t -> data, p, &ln_determinant[j]);
} catch (MelderError) {
// Try the alternative: the pooled covariance matrix.
// Clear the error.
Melder_clearError ();
if (npool == 0) {
NUMlowerCholeskyInverse (pool -> data, p, &lnd);
}
npool++;
sscpvec[j] = pool.peek();
ln_determinant[j] = lnd;
}
}
if (npool > 0) {
Melder_warning (npool, U" groups use pooled covariance matrix.");
}
}
// Labels for columns in ClassificationTable
for (long j = 1; j <= g; j++) {
const char32 *name = Thing_getName ( (Thing) my groups -> item[j]);
if (! name) {
name = U"?";
}
TableOfReal_setColumnLabel (him.peek(), j, name);
}
// Normalize the sum of the apriori probabilities to 1.
// Next take ln (p) because otherwise probabilities might be too small to represent.
NUMvector_normalize1 (my aprioriProbabilities, g);
double logg = log (g);
for (long j = 1; j <= g; j++) {
log_apriori[j] = useAprioriProbabilities ? log (my aprioriProbabilities[j]) : - logg;
}
// Generalized squared distance function:
// D^2(x) = (x - mu)' S^-1 (x - mu) + ln (determinant(S)) - 2 ln (apriori)
for (long i = 1; i <= m; i++) {
double norm = 0, pt_max = -1e38;
for (long j = 1; j <= g; j++) {
SSCP t = (SSCP) groups -> item[j];
double md = mahalanobisDistanceSq (sscpvec[j] -> data, p, thy data[i], t -> centroid, buf.peek());
double pt = log_apriori[j] - 0.5 * (ln_determinant[j] + md);
if (pt > pt_max) {
pt_max = pt;
}
log_p[j] = pt;
}
for (long j = 1; j <= g; j++) {
norm += (log_p[j] = exp (log_p[j] - pt_max));
}
for (long j = 1; j <= g; j++) {
his data[i][j] = log_p[j] / norm;
}
}
return him.transfer();
} catch (MelderError) {
Melder_throw (U"ClassificationTable from Discriminant & TableOfReal not created.");
}
}
