PProbabilityEstimator TProbabilityEstimatorConstructor_m::operator()(PDistribution frequencies, PDistribution apriori, PExampleGenerator, const long &weightID, const int &) const
{ TProbabilityEstimator_FromDistribution *pefd = mlnew TProbabilityEstimator_FromDistribution(CLONE(TDistribution, frequencies));
PProbabilityEstimator estimator = pefd;
TDiscDistribution *ddist = pefd->probabilities.AS(TDiscDistribution);
if (ddist && (ddist->cases > 1e-20) && apriori) {
TDiscDistribution *dapriori = apriori.AS(TDiscDistribution);
if (!dapriori || (dapriori->abs < 1e-20))
raiseError("invalid apriori distribution");
float mabs = m/dapriori->abs;
const float &abs = ddist->abs;
const float &cases = ddist->cases;
const float div = cases + m;
if ((abs==cases) || !renormalize) {
int i = 0;
for(TDiscDistribution::iterator di(ddist->begin()), de(ddist->end()), ai(dapriori->begin());
di != de;
di++, ai++, i++)
ddist->setint(i, (*di+*ai*mabs)/div);
}
else {
int i = 0;
for(TDiscDistribution::iterator di(ddist->begin()), de(ddist->end()), ai(dapriori->begin());
di != de;
di++, ai++, i++)
ddist->setint(i, (*di / abs * cases + *ai*mabs)/div);
}
}
else
pefd->probabilities->normalize();
return estimator;
}
PProbabilityEstimator TProbabilityEstimatorConstructor_Laplace::operator()(PDistribution frequencies, PDistribution, PExampleGenerator, const long &, const int &) const
{ TProbabilityEstimator_FromDistribution *pefd = mlnew TProbabilityEstimator_FromDistribution(CLONE(TDistribution, frequencies));
PProbabilityEstimator estimator = pefd;
TDiscDistribution *ddist = pefd->probabilities.AS(TDiscDistribution);
if (ddist) {
const float &abs = ddist->abs;
const float &cases = ddist->cases;
const float div = cases + l * ddist->noOfElements();
int i = 0;
if (div) {
if ((cases == abs) || !renormalize || (abs<1e-20))
PITERATE(TDiscDistribution, di, ddist)
ddist->setint(i++, (*di + l) / div);
else
PITERATE(TDiscDistribution, di, ddist)
ddist->setint(i++, (*di / abs * cases + l) / div);
}
else
pefd->probabilities->normalize();
}
else
pefd->probabilities->normalize();
return estimator;
}
// rejects the split if there are less than two non-empty branches
// or there is a non-empty branch with less then minSubset examples
bool checkDistribution(const TDiscDistribution &dist, const float &minSubset)
{
int nonzero = 0;
for(TDiscDistribution::const_iterator dvi(dist.begin()), dve(dist.end()); dvi!=dve; dvi++)
if (*dvi > 0) {
if (*dvi < minSubset)
return false;
nonzero++;
}
return nonzero >= 2;
}
PConditionalProbabilityEstimator TConditionalProbabilityEstimatorConstructor_loess::operator()(PContingency frequencies, PDistribution, PExampleGenerator, const long &, const int &) const
{ if (frequencies->varType != TValue::FLOATVAR)
if (frequencies->outerVariable)
raiseError("attribute '%s' is not continuous", frequencies->outerVariable->get_name().c_str());
else
raiseError("continuous attribute expected for condition");
if (!frequencies->continuous->size())
// This is ugly, but: if you change this, you should also change the code which catches it in
// Bayesian learner
raiseError("distribution (of attribute values, probably) is empty or has only a single value");
PContingency cont = CLONE(TContingency, frequencies);
const TDistributionMap &points = *frequencies->continuous;
/* if (frequencies->continuous->size() == 1) {
TDiscDistribution *f = (TDiscDistribution *)(points.begin()->second.getUnwrappedPtr());
f->normalize();
f->variances = mlnew TFloatList(f->size(), 0.0);
return mlnew TConditionalProbabilityEstimator_FromDistribution(cont);
}
*/
cont->continuous->clear();
vector<float> xpoints;
distributePoints(points, nPoints, xpoints, distributionMethod);
if (!xpoints.size())
raiseError("no points for the curve (check 'nPoints')");
if (frequencies->continuous->size() == 1) {
TDiscDistribution *f = (TDiscDistribution *)(points.begin()->second.getUnwrappedPtr());
f->normalize();
f->variances = mlnew TFloatList(f->size(), 0.0);
const_ITERATE(vector<float>, pi, xpoints)
(*cont->continuous)[*pi] = f;
return mlnew TConditionalProbabilityEstimator_FromDistribution(cont);
}
TDistributionMap::const_iterator lowedge = points.begin();
TDistributionMap::const_iterator highedge = points.end();
bool needAll;
map<float, PDistribution>::const_iterator from, to;
vector<float>::const_iterator pi(xpoints.begin()), pe(xpoints.end());
float refx = *pi;
from = lowedge;
to = highedge;
int totalNumOfPoints = frequencies->outerDistribution->abs;
int needpoints = int(ceil(totalNumOfPoints * windowProportion));
if (needpoints<3)
needpoints = 3;
TSimpleRandomGenerator rgen(frequencies->outerDistribution->cases);
if ((needpoints<=0) || (needpoints>=totalNumOfPoints)) { //points.size()
needAll = true;
from = lowedge;
to = highedge;
}
else {
needAll = false;
/* Find the window */
from = points.lower_bound(refx);
to = points.upper_bound(refx);
if (from==to)
if (to != highedge)
to++;
else
from --;
/* Extend the interval; we set from to highedge when it would go beyond lowedge, to indicate that only to can be modified now */
while (needpoints > 0) {
if ((to == highedge) || ((from != highedge) && (refx - (*from).first < (*to).first - refx))) {
if (from == lowedge)
from = highedge;
else {
from--;
needpoints -= (*from).second->cases;
}
}
else {
to++;
if (to!=highedge)
needpoints -= (*to).second->cases;
else
needpoints = 0;
}
}
if (from == highedge)
from = lowedge;
/* else
from++;*/
}
int numOfOverflowing = 0;
// This follows http://www-2.cs.cmu.edu/afs/cs/project/jair/pub/volume4/cohn96a-html/node7.html
for(;;) {
TDistributionMap::const_iterator tt = to;
--tt;
if (tt == from) {
TDistribution *Sy = CLONE(TDistribution, (*tt).second);
PDistribution wSy = Sy;
Sy->normalize();
(*cont->continuous)[refx] = (wSy);
((TDiscDistribution *)(Sy)) ->variances = mlnew TFloatList(Sy->variable->noOfValues(), 0.0);
}
else {
float h = (refx - (*from).first);
if ((*tt).first - refx > h)
h = ((*tt).first - refx);
/* Iterate through the window */
tt = from;
const float &x = (*tt).first;
const PDistribution &y = (*tt).second;
float cases = y->abs;
float w = fabs(refx - x) / h;
w = 1 - w*w*w;
w = w*w*w;
const float num = y->abs; // number of instances with this x - value
float n = w * num;
float Sww = w * w * num;
float Sx = w * x * num;
float Swwx = w * w * x * num;
float Swwxx = w * w * x * x * num;
TDistribution *Sy = CLONE(TDistribution, y);
PDistribution wSy = Sy;
*Sy *= w;
float Sxx = w * x * x * num;
TDistribution *Syy = CLONE(TDistribution, y);
PDistribution wSyy = Syy;
*Syy *= w;
TDistribution *Sxy = CLONE(TDistribution, y);
PDistribution wSxy = Sxy;
*Sxy *= w * x;
if (tt!=to)
while (++tt != to) {
const float &x = (*tt).first;
const PDistribution &y = (*tt).second;
cases += y->abs;
w = fabs(refx - x) / h;
w = 1 - w*w*w;
w = w*w*w;
const float num = y->abs;
n += w * num;
Sww += w * w * num;
Sx += w * x * num;
Swwx += w * w * x * num;
Swwxx += w * w * x * x * num;
Sxx += w * x * x * num;
TDistribution *ty = CLONE(TDistribution, y);
PDistribution wty = ty;
*ty *= w;
*Sy += wty;
*Syy += wty;
*ty *= x;
*Sxy += wty;
//*ty *= PDistribution(y);
}
float sigma_x2 = n<1e-6 ? 0.0 : (Sxx - Sx * Sx / n)/n;
if (sigma_x2<1e-10) {
*Sy *= 0;
Sy->cases = cases;
(*cont->continuous)[refx] = (wSy);
}
TDistribution *sigma_y2 = CLONE(TDistribution, Sy);
PDistribution wsigma_y2 = sigma_y2;
*sigma_y2 *= wsigma_y2;
*sigma_y2 *= -1/n;
*sigma_y2 += wSyy;
*sigma_y2 *= 1/n;
TDistribution *sigma_xy = CLONE(TDistribution, Sy);
PDistribution wsigma_xy = sigma_xy;
*sigma_xy *= -Sx/n;
*sigma_xy += wSxy;
*sigma_xy *= 1/n;
// This will be sigma_xy / sigma_x2, but we'll multiply it by whatever we need
TDistribution *sigma_tmp = CLONE(TDistribution, sigma_xy);
PDistribution wsigma_tmp = sigma_tmp;
//*sigma_tmp *= wsigma_tmp;
if (sigma_x2 > 1e-10)
*sigma_tmp *= 1/sigma_x2;
const float difx = refx - Sx/n;
// computation of y
*sigma_tmp *= difx;
*Sy *= 1/n;
*Sy += *sigma_tmp;
// probabilities that are higher than 0.9 normalize with a logistic function, which produces two positive
// effects: prevents overfitting and avoids probabilities that are higher than 1.0. But, on the other hand, this
// solution is rather unmathematical. Do the same for probabilities that are lower than 0.1.
vector<float>::iterator syi(((TDiscDistribution *)(Sy))->distribution.begin());
vector<float>::iterator sye(((TDiscDistribution *)(Sy))->distribution.end());
for (; syi!=sye; syi++) {
if (*syi > 0.9) {
Sy->abs -= *syi;
*syi = 1/(1+exp(-10*((*syi)-0.9)*log(9.0)-log(9.0)));
Sy->abs += *syi;
}
if (*syi < 0.1) {
Sy->abs -= *syi;
*syi = 1/(1+exp(10*(0.1-(*syi))*log(9.0)+log(9.0)));
Sy->abs += *syi;
}
}
Sy->cases = cases;
Sy->normalize();
(*cont->continuous)[refx] = (wSy);
// now for the variance
// restore sigma_tmp and compute the conditional sigma
if ((fabs(difx) > 1e-10) && (sigma_x2 > 1e-10)) {
*sigma_tmp *= (1/difx);
*sigma_tmp *= wsigma_xy;
*sigma_tmp *= -1;
*sigma_tmp += wsigma_y2;
// fct corresponds to part of (10) in the brackets (see URL above)
// float fct = Sww + difx*difx/sigma_x2/sigma_x2 * (Swwxx - 2/n * Sx*Swwx + 2/n/n * Sx*Sx*Sww);
float fct = 1 + difx*difx/sigma_x2; //n + difx*difx/sigma_x2+n*n --- add this product to the overall fct sum if you are estimating error for a single user and not for the line.
*sigma_tmp *= fct/n; // fct/n/n;
}
((TDiscDistribution *)(Sy)) ->variances = mlnew TFloatList(((TDiscDistribution *)(sigma_tmp))->distribution);
}
// on to the next point
pi++;
if (pi==pe)
break;
refx = *pi;
// Adjust the window
while (to!=highedge) {
float dif = (refx - (*from).first) - ((*to).first - refx);
if ((dif>0) || (dif==0) && rgen.randbool()) {
if (numOfOverflowing > 0) {
from++;
numOfOverflowing -= (*from).second->cases;
}
else {
to++;
if (to!=highedge)
numOfOverflowing += (*to).second->cases;
}
}
else
break;
}
}
return mlnew TConditionalProbabilityEstimator_FromDistribution(cont);
}
PDistribution TLogRegClassifier::classDistribution(const TExample &origexam)
{
checkProperty(domain);
TExample cexample(domain, origexam);
TExample *example2;
if (imputer)
example2 = imputer->call(cexample);
else {
if (dataDescription)
for(TExample::const_iterator ei(cexample.begin()), ee(cexample.end()-1); ei!=ee; ei++)
if ((*ei).isSpecial())
return TClassifier::classDistribution(cexample, dataDescription);
example2 = &cexample;
}
TExample *example = continuizedDomain ? mlnew TExample(continuizedDomain, *example2) : example2;
float prob1;
try {
// multiply example with beta
TAttributedFloatList::const_iterator b(beta->begin()), be(beta->end());
// get beta 0
prob1 = *b;
b++;
// multiply beta with example
TVarList::const_iterator vi(example->domain->attributes->begin());
TExample::const_iterator ei(example->begin()), ee(example->end());
for (; (b!=be) && (ei!=ee); ei++, b++, vi++) {
if ((*ei).isSpecial())
raiseError("unknown value in attribute '%s'", (*vi)->get_name().c_str());
prob1 += (*ei).floatV * (*b);
}
prob1 = exp(prob1)/(1+exp(prob1));
}
catch (...) {
if (imputer)
mldelete example2;
if (continuizedDomain)
mldelete example;
throw;
}
if (imputer)
mldelete example2;
if (continuizedDomain)
mldelete example;
if (classVar->varType == TValue::INTVAR) {
TDiscDistribution *dist = mlnew TDiscDistribution(classVar);
PDistribution res = dist;
dist->addint(0, 1-prob1);
dist->addint(1, prob1);
return res;
}
else {
TContDistribution *dist = mlnew TContDistribution(classVar);
PDistribution res = dist;
dist->addfloat(prob1, 1.0);
return res;
}
}
