void C_surrogates(SEXP node, SEXP learnsample, SEXP weights, SEXP controls,
SEXP fitmem) {
SEXP x, y, expcovinf;
SEXP splitctrl, inputs;
SEXP split, thiswhichNA;
int nobs, ninputs, i, j, k, jselect, maxsurr, *order, nvar = 0;
double ms, cp, *thisweights, *cutpoint, *maxstat,
*splitstat, *dweights, *tweights, *dx, *dy;
double cut, *twotab, *ytmp, sumw = 0.0;
nobs = get_nobs(learnsample);
ninputs = get_ninputs(learnsample);
splitctrl = get_splitctrl(controls);
maxsurr = get_maxsurrogate(splitctrl);
inputs = GET_SLOT(learnsample, PL2_inputsSym);
jselect = S3get_variableID(S3get_primarysplit(node));
/* (weights > 0) in left node are the new `response' to be approximated */
y = S3get_nodeweights(VECTOR_ELT(node, S3_LEFT));
ytmp = Calloc(nobs, double);
for (i = 0; i < nobs; i++) {
ytmp[i] = REAL(y)[i];
if (ytmp[i] > 1.0) ytmp[i] = 1.0;
}
for (j = 0; j < ninputs; j++) {
if (is_nominal(inputs, j + 1)) continue;
nvar++;
}
nvar--;
if (maxsurr != LENGTH(S3get_surrogatesplits(node)))
error("nodes does not have %d surrogate splits", maxsurr);
if (maxsurr > nvar)
error("cannot set up %d surrogate splits with only %d ordered input variable(s)",
maxsurr, nvar);
tweights = Calloc(nobs, double);
dweights = REAL(weights);
for (i = 0; i < nobs; i++) tweights[i] = dweights[i];
if (has_missings(inputs, jselect)) {
thiswhichNA = get_missings(inputs, jselect);
for (k = 0; k < LENGTH(thiswhichNA); k++)
tweights[INTEGER(thiswhichNA)[k] - 1] = 0.0;
}
/* check if sum(weights) > 1 */
sumw = 0.0;
for (i = 0; i < nobs; i++) sumw += tweights[i];
if (sumw < 2.0)
error("can't implement surrogate splits, not enough observations available");
expcovinf = GET_SLOT(fitmem, PL2_expcovinfssSym);
C_ExpectCovarInfluence(ytmp, 1, tweights, nobs, expcovinf);
splitstat = REAL(get_splitstatistics(fitmem));
/* <FIXME> extend `TreeFitMemory' to those as well ... */
maxstat = Calloc(ninputs, double);
cutpoint = Calloc(ninputs, double);
order = Calloc(ninputs, int);
/* <FIXME> */
/* this is essentially an exhaustive search */
/* <FIXME>: we don't want to do this for random forest like trees
</FIXME>
*/
for (j = 0; j < ninputs; j++) {
order[j] = j + 1;
maxstat[j] = 0.0;
cutpoint[j] = 0.0;
/* ordered input variables only (for the moment) */
if ((j + 1) == jselect || is_nominal(inputs, j + 1))
continue;
x = get_variable(inputs, j + 1);
if (has_missings(inputs, j + 1)) {
/* update _tweights_ wrt missings in variable j + 1 */
thisweights = C_tempweights(j + 1, tweights, fitmem, inputs);
/* check if sum(weights) > 1 */
sumw = 0.0;
for (i = 0; i < nobs; i++) sumw += thisweights[i];
if (sumw < 2.0) continue;
C_ExpectCovarInfluence(ytmp, 1, thisweights, nobs, expcovinf);
C_split(REAL(x), 1, ytmp, 1, thisweights, nobs,
INTEGER(get_ordering(inputs, j + 1)), splitctrl,
GET_SLOT(fitmem, PL2_linexpcov2sampleSym),
expcovinf, 1, &cp, &ms, splitstat);
} else {
C_split(REAL(x), 1, ytmp, 1, tweights, nobs,
INTEGER(get_ordering(inputs, j + 1)), splitctrl,
GET_SLOT(fitmem, PL2_linexpcov2sampleSym),
expcovinf, 1, &cp, &ms, splitstat);
}
maxstat[j] = -ms;
cutpoint[j] = cp;
}
/* order with respect to maximal statistic */
rsort_with_index(maxstat, order, ninputs);
twotab = Calloc(4, double);
/* the best `maxsurr' ones are implemented */
for (j = 0; j < maxsurr; j++) {
if (is_nominal(inputs, order[j])) continue;
for (i = 0; i < 4; i++) twotab[i] = 0.0;
cut = cutpoint[order[j] - 1];
/* this might give warnings about split being
UNPROTECTed but is is since node is PROTECTed */
PROTECT(split = allocVector(VECSXP, SPLIT_LENGTH));
SET_VECTOR_ELT(S3get_surrogatesplits(node), j, split);
C_init_orderedsplit(split, 0);
S3set_variableID(split, order[j]);
REAL(S3get_splitpoint(split))[0] = cut;
dx = REAL(get_variable(inputs, order[j]));
dy = REAL(y);
/* OK, this is a dirty hack: determine if the split
goes left or right by the Pearson residual of a 2x2 table.
I don't want to use the big caliber here
*/
for (i = 0; i < nobs; i++) {
twotab[0] += ((dy[i] == 1) && (dx[i] <= cut)) * tweights[i];
twotab[1] += (dy[i] == 1) * tweights[i];
twotab[2] += (dx[i] <= cut) * tweights[i];
twotab[3] += tweights[i];
}
S3set_toleft(split, (int) (twotab[0] - twotab[1] * twotab[2] /
twotab[3]) > 0);
UNPROTECT(1);
}
Free(maxstat);
Free(cutpoint);
Free(order);
Free(tweights);
Free(twotab);
Free(ytmp);
}
void C_Node(SEXP node, SEXP learnsample, SEXP weights,
SEXP fitmem, SEXP controls, int TERMINAL, int depth) {
int nobs, ninputs, jselect, q, j, k, i;
double mincriterion, sweights, *dprediction;
double *teststat, *pvalue, smax, cutpoint = 0.0, maxstat = 0.0;
double *standstat, *splitstat;
SEXP responses, inputs, x, expcovinf, linexpcov;
SEXP varctrl, splitctrl, gtctrl, tgctrl, split, testy, predy;
double *dxtransf, *thisweights;
int *itable;
nobs = get_nobs(learnsample);
ninputs = get_ninputs(learnsample);
varctrl = get_varctrl(controls);
splitctrl = get_splitctrl(controls);
gtctrl = get_gtctrl(controls);
tgctrl = get_tgctrl(controls);
mincriterion = get_mincriterion(gtctrl);
responses = GET_SLOT(learnsample, PL2_responsesSym);
inputs = GET_SLOT(learnsample, PL2_inputsSym);
testy = get_test_trafo(responses);
predy = get_predict_trafo(responses);
q = ncol(testy);
/* <FIXME> we compute C_GlobalTest even for TERMINAL nodes! </FIXME> */
/* compute the test statistics and the node criteria for each input */
C_GlobalTest(learnsample, weights, fitmem, varctrl,
gtctrl, get_minsplit(splitctrl),
REAL(S3get_teststat(node)), REAL(S3get_criterion(node)), depth);
/* sum of weights: C_GlobalTest did nothing if sweights < mincriterion */
sweights = REAL(GET_SLOT(GET_SLOT(fitmem, PL2_expcovinfSym),
PL2_sumweightsSym))[0];
REAL(VECTOR_ELT(node, S3_SUMWEIGHTS))[0] = sweights;
/* compute the prediction of this node */
dprediction = REAL(S3get_prediction(node));
/* <FIXME> feed raw numeric values OR dummy encoded factors as y
Problem: what happens for survival times ? */
C_prediction(REAL(predy), nobs, ncol(predy), REAL(weights),
sweights, dprediction);
/* </FIXME> */
teststat = REAL(S3get_teststat(node));
pvalue = REAL(S3get_criterion(node));
/* try the two out of ninputs best inputs variables */
/* <FIXME> be more flexible and add a parameter controlling
the number of inputs tried </FIXME> */
for (j = 0; j < 2; j++) {
smax = C_max(pvalue, ninputs);
REAL(S3get_maxcriterion(node))[0] = smax;
/* if the global null hypothesis was rejected */
if (smax > mincriterion && !TERMINAL) {
/* the input variable with largest association to the response */
jselect = C_whichmax(pvalue, teststat, ninputs) + 1;
/* get the raw numeric values or the codings of a factor */
x = get_variable(inputs, jselect);
if (has_missings(inputs, jselect)) {
expcovinf = GET_SLOT(get_varmemory(fitmem, jselect),
PL2_expcovinfSym);
thisweights = C_tempweights(jselect, weights, fitmem, inputs);
} else {
expcovinf = GET_SLOT(fitmem, PL2_expcovinfSym);
thisweights = REAL(weights);
}
/* <FIXME> handle ordered factors separatly??? </FIXME> */
if (!is_nominal(inputs, jselect)) {
/* search for a split in a ordered variable x */
split = S3get_primarysplit(node);
/* check if the n-vector of splitstatistics
should be returned for each primary split */
if (get_savesplitstats(tgctrl)) {
C_init_orderedsplit(split, nobs);
splitstat = REAL(S3get_splitstatistics(split));
} else {
C_init_orderedsplit(split, 0);
splitstat = REAL(get_splitstatistics(fitmem));
}
C_split(REAL(x), 1, REAL(testy), q, thisweights, nobs,
INTEGER(get_ordering(inputs, jselect)), splitctrl,
GET_SLOT(fitmem, PL2_linexpcov2sampleSym),
expcovinf, REAL(S3get_splitpoint(split)), &maxstat,
splitstat);
S3set_variableID(split, jselect);
} else {
/* search of a set of levels (split) in a numeric variable x */
split = S3get_primarysplit(node);
/* check if the n-vector of splitstatistics
should be returned for each primary split */
if (get_savesplitstats(tgctrl)) {
C_init_nominalsplit(split,
LENGTH(get_levels(inputs, jselect)),
nobs);
splitstat = REAL(S3get_splitstatistics(split));
} else {
C_init_nominalsplit(split,
LENGTH(get_levels(inputs, jselect)),
0);
splitstat = REAL(get_splitstatistics(fitmem));
}
linexpcov = get_varmemory(fitmem, jselect);
standstat = Calloc(get_dimension(linexpcov), double);
C_standardize(REAL(GET_SLOT(linexpcov,
PL2_linearstatisticSym)),
REAL(GET_SLOT(linexpcov, PL2_expectationSym)),
REAL(GET_SLOT(linexpcov, PL2_covarianceSym)),
get_dimension(linexpcov), get_tol(splitctrl),
standstat);
C_splitcategorical(INTEGER(x),
LENGTH(get_levels(inputs, jselect)),
REAL(testy), q, thisweights,
nobs, standstat, splitctrl,
GET_SLOT(fitmem, PL2_linexpcov2sampleSym),
expcovinf, &cutpoint,
INTEGER(S3get_splitpoint(split)),
&maxstat, splitstat);
/* compute which levels of a factor are available in this node
(for printing) later on. A real `table' for this node would
induce too much overhead here. Maybe later. */
itable = INTEGER(S3get_table(split));
dxtransf = REAL(get_transformation(inputs, jselect));
for (k = 0; k < LENGTH(get_levels(inputs, jselect)); k++) {
itable[k] = 0;
for (i = 0; i < nobs; i++) {
if (dxtransf[k * nobs + i] * thisweights[i] > 0) {
itable[k] = 1;
continue;
}
}
}
Free(standstat);
}
if (maxstat == 0) {
if (j == 1) {
S3set_nodeterminal(node);
} else {
/* do not look at jselect in next iteration */
pvalue[jselect - 1] = R_NegInf;
}
} else {
S3set_variableID(split, jselect);
break;
}
} else {