// Stopping criteria
boolean isDone(KMeansModel model, double[][] newCenters, double[][] oldCenters) {
if (!isRunning()) return true; // Stopped/cancelled
// Stopped for running out iterations
if (model._output._iterations >= _parms._max_iterations) return true;
// Compute average change in standardized cluster centers
if (oldCenters == null) return false; // No prior iteration, not stopping
double average_change = 0;
for (int clu = 0; clu < _parms._k; clu++)
average_change +=
hex.genmodel.GenModel.KMeans_distance(
oldCenters[clu], newCenters[clu], _isCats, null, null);
average_change /= _parms._k; // Average change per cluster
model._output._avg_centroids_chg =
ArrayUtils.copyAndFillOf(
model._output._avg_centroids_chg,
model._output._avg_centroids_chg.length + 1,
average_change);
model._output._training_time_ms =
ArrayUtils.copyAndFillOf(
model._output._training_time_ms,
model._output._training_time_ms.length + 1,
System.currentTimeMillis());
return average_change < TOLERANCE;
}
/**
* Bulk scoring API for one row. Chunks are all compatible with the model, and expect the last
* Chunks are for the final distribution and prediction. Default method is to just load the data
* into the tmp array, then call subclass scoring logic.
*/
@Override
protected double[] score0(
double data[ /*ncols*/], double preds[ /*nclasses+1*/], double weight, double offset) {
super.score0(data, preds, weight, offset); // These are f_k(x) in Algorithm 10.4
if (_parms._distribution == Distributions.Family.bernoulli) {
double f =
preds[1]
+ _output._init_f
+ offset; // Note: class 1 probability stored in preds[1] (since we have only one
// tree)
preds[2] = _parms._distribution.linkInv(f);
preds[1] = 1.0 - preds[2];
} else if (_parms._distribution
== Distributions.Family.multinomial) { // Kept the initial prediction for binomial
if (_output.nclasses() == 2) { // 1-tree optimization for binomial
preds[1] +=
_output._init_f
+ offset; // offset is not yet allowed, but added here to be future-proof
preds[2] = -preds[1];
}
hex.genmodel.GenModel.GBM_rescale(preds);
} else { // Regression
double f = preds[0] + _output._init_f + offset;
preds[0] = _parms._distribution.linkInv(f);
}
return preds;
}
// Call builder specific score code and then correct probabilities
// if it is necessary.
void score2(Chunk chks[], double weight, double offset, double fs[ /*nclass*/], int row) {
double sum = score1(chks, weight, offset, fs, row);
if (isClassifier()) {
if (!Double.isInfinite(sum) && sum > 0f && sum != 1f) ArrayUtils.div(fs, sum);
if (_parms._balance_classes)
GenModel.correctProbabilities(
fs, _model._output._priorClassDist, _model._output._modelClassDist);
}
}
@Override
public void map(Chunk[] cs) {
for (int row = 0; row < cs[0]._len; row++) {
double[] values = new double[cs.length];
// fetch the data - using consistent NA and categorical data handling (same as for training)
data(values, cs, row, _means, _mults, _modes);
// compute the distance from the (standardized) cluster centroids
_tss += hex.genmodel.GenModel.KMeans_distance(_gc, values, _isCats, null, null);
}
}
/** Return both nearest of N cluster center/centroids, and the square-distance. */
private static ClusterDist closest(
double[][] centers, double[] point, String[][] isCats, ClusterDist cd, int count) {
int min = -1;
double minSqr = Double.MAX_VALUE;
for (int cluster = 0; cluster < count; cluster++) {
double sqr =
hex.genmodel.GenModel.KMeans_distance(centers[cluster], point, isCats, null, null);
if (sqr < minSqr) { // Record nearest cluster
min = cluster;
minSqr = sqr;
}
}
cd._cluster = min; // Record nearest cluster
cd._dist = minSqr; // Record square-distance
return cd; // Return for flow-coding
}
// Note: For small probabilities, product may end up zero due to underflow error. Can circumvent
// by taking logs.
@Override
protected double[] score0(double[] data, double[] preds) {
double[] nums = new double[_output._levels.length]; // log(p(x,y)) for all levels of y
assert preds.length
== (_output._levels.length + 1); // Note: First column of preds is predicted response class
// Compute joint probability of predictors for every response class
for (int rlevel = 0; rlevel < _output._levels.length; rlevel++) {
// Take logs to avoid overflow: p(x,y) = p(x|y)*p(y) -> log(p(x,y)) = log(p(x|y)) + log(p(y))
nums[rlevel] = Math.log(_output._apriori_raw[rlevel]);
for (int col = 0; col < _output._ncats; col++) {
if (Double.isNaN(data[col])) continue; // Skip predictor in joint x_1,...,x_m if NA
int plevel = (int) data[col];
double prob =
plevel < _output._pcond_raw.length
? _output._pcond_raw[col][rlevel][plevel]
: _parms._laplace
/ ((double) _output._rescnt[rlevel]
+ _parms._laplace
* _output
._domains[col]
.length); // Laplace smoothing if predictor level unobserved in
// training set
nums[rlevel] +=
Math.log(
prob <= _parms._eps_prob
? _parms._min_prob
: prob); // log(p(x|y)) = \sum_{j = 1}^m p(x_j|y)
}
// For numeric predictors, assume Gaussian distribution with sample mean and variance from
// model
for (int col = _output._ncats; col < data.length; col++) {
if (Double.isNaN(data[col])) continue; // Skip predictor in joint x_1,...,x_m if NA
double x = data[col];
double mean =
Double.isNaN(_output._pcond_raw[col][rlevel][0])
? 0
: _output._pcond_raw[col][rlevel][0];
double stddev =
Double.isNaN(_output._pcond_raw[col][rlevel][1])
? 1.0
: (_output._pcond_raw[col][rlevel][1] <= _parms._eps_sdev
? _parms._min_sdev
: _output._pcond_raw[col][rlevel][1]);
// double prob = Math.exp(new NormalDistribution(mean, stddev).density(data[col])); //
// slower
double prob =
Math.exp(-((x - mean) * (x - mean)) / (2. * stddev * stddev))
/ (stddev * Math.sqrt(2. * Math.PI)); // faster
nums[rlevel] += Math.log(prob <= _parms._eps_prob ? _parms._min_prob : prob);
}
}
// Numerically unstable:
// p(x,y) = exp(log(p(x,y))), p(x) = \Sum_{r = levels of y} exp(log(p(x,y = r))) -> p(y|x) =
// p(x,y)/p(x)
// Instead, we rewrite using a more stable form:
// p(y|x) = p(x,y)/p(x) = exp(log(p(x,y))) / (\Sum_{r = levels of y} exp(log(p(x,y = r)))
// = 1 / ( exp(-log(p(x,y))) * \Sum_{r = levels of y} exp(log(p(x,y = r))) )
// = 1 / ( \Sum_{r = levels of y} exp( log(p(x,y = r)) - log(p(x,y)) ))
for (int i = 0; i < nums.length; i++) {
double sum = 0;
for (int j = 0; j < nums.length; j++) sum += Math.exp(nums[j] - nums[i]);
preds[i + 1] = 1 / sum;
}
// Select class with highest conditional probability
preds[0] = GenModel.getPrediction(preds, _output._priorClassDist, data, defaultThreshold());
return preds;
}