/**
* Builds L2-regularized regressors for a sequence of regularization parameter lambdas on sparse
* inputs. Each row of the input represents a feature (instead of a data point), i.e., in
* column-oriented format. This procedure does not assume the data is normalized or centered.
*
* @param attrs the attribute list.
* @param indices the indices.
* @param values the values.
* @param y the targets.
* @param maxNumIters the maximum number of iterations.
* @param lambdas the lambdas array.
* @return L2-regularized regressors.
*/
public GLM[] buildGaussianRegressors(
int[] attrs,
int[][] indices,
double[][] values,
double[] y,
int maxNumIters,
double[] lambdas) {
double[] w = new double[attrs.length];
double intercept = 0;
GLM[] glms = new GLM[lambdas.length];
Arrays.sort(lambdas);
// Backup targets
double[] rTrain = new double[y.length];
for (int i = 0; i < rTrain.length; i++) {
rTrain[i] = y[i];
}
// Calculate sum of squares
double[] sq = new double[attrs.length];
for (int i = 0; i < values.length; i++) {
sq[i] = StatUtils.sumSq(values[i]);
}
// Compute the regularization path
for (int g = 0; g < glms.length; g++) {
double lambda = lambdas[g];
// Coordinate descent
final double tl2 = lambda * y.length;
for (int iter = 0; iter < maxNumIters; iter++) {
double prevLoss = GLMOptimUtils.computeRidgeLoss(rTrain, w, lambda);
if (fitIntercept) {
intercept += OptimUtils.fitIntercept(rTrain);
}
doOnePassGaussian(indices, values, sq, tl2, w, rTrain);
double currLoss = GLMOptimUtils.computeRidgeLoss(rTrain, w, lambda);
if (verbose) {
System.out.println("Iteration " + iter + ": " + " " + currLoss);
}
if (OptimUtils.isConverged(prevLoss, currLoss, epsilon)) {
break;
}
}
glms[g] = GLMOptimUtils.getGLM(attrs, w, intercept, LinkFunction.IDENTITY);
}
return glms;
}
/**
* Builds L2-regularized binary classifiers for a sequence of regularization parameter lambdas on
* sparse format. Each row of the input represents a feature (instead of a data point), i.e., in
* column-oriented format. This procedure does not assume the data is normalized or centered.
*
* @param attrs the attribute list.
* @param indices the indices.
* @param values the values.
* @param y the targets.
* @param maxNumIters the maximum number of iterations.
* @param lambdas the lambdas array.
* @return L2-regularized classifiers.
*/
public GLM[] buildBinaryClassifiers(
int[] attrs,
int[][] indices,
double[][] values,
double[] y,
int maxNumIters,
double[] lambdas) {
double[] w = new double[attrs.length];
double intercept = 0;
double[] pTrain = new double[y.length];
double[] rTrain = new double[y.length];
OptimUtils.computePseudoResidual(pTrain, y, rTrain);
// Calculate theta's
double[] theta = new double[values.length];
for (int i = 0; i < values.length; i++) {
theta[i] = StatUtils.sumSq(values[i]) / 4;
}
GLM[] glms = new GLM[lambdas.length];
Arrays.sort(lambdas);
for (int g = 0; g < glms.length; g++) {
double lambda = lambdas[g];
// Coordinate gradient descent
final double tl2 = lambda * y.length;
for (int iter = 0; iter < maxNumIters; iter++) {
double prevLoss = GLMOptimUtils.computeRidgeLoss(pTrain, y, w, lambda);
if (fitIntercept) {
intercept += OptimUtils.fitIntercept(pTrain, rTrain, y);
}
doOnePassBinomial(indices, values, theta, y, tl2, w, pTrain, rTrain);
double currLoss = GLMOptimUtils.computeRidgeLoss(pTrain, y, w, lambda);
if (verbose) {
System.out.println("Iteration " + iter + ": " + " " + currLoss);
}
if (OptimUtils.isConverged(prevLoss, currLoss, epsilon)) {
break;
}
}
glms[g] = GLMOptimUtils.getGLM(attrs, w, intercept, LinkFunction.LOGIT);
}
return glms;
}
/**
* Builds an L2-regularized regressor on sparse inputs. Each row of the input represents a feature
* (instead of a data point), i.e., in column-oriented format. This procedure does not assume the
* data is normalized or centered.
*
* @param attrs the attribute list.
* @param indices the indices.
* @param values the values.
* @param y the targets.
* @param maxNumIters the maximum number of iterations.
* @param lambda the lambda.
* @return an L2-regularized regressor.
*/
public GLM buildGaussianRegressor(
int[] attrs, int[][] indices, double[][] values, double[] y, int maxNumIters, double lambda) {
double[] w = new double[attrs.length];
double intercept = 0;
// Initialize residuals
double[] rTrain = new double[y.length];
for (int i = 0; i < rTrain.length; i++) {
rTrain[i] = y[i];
}
// Calculate sum of squares
double[] sq = new double[attrs.length];
for (int i = 0; i < values.length; i++) {
sq[i] = StatUtils.sumSq(values[i]);
}
// Coordinate descent
final double tl2 = lambda * y.length;
for (int iter = 0; iter < maxNumIters; iter++) {
double prevLoss = GLMOptimUtils.computeRidgeLoss(rTrain, w, lambda);
if (fitIntercept) {
intercept += OptimUtils.fitIntercept(rTrain);
}
doOnePassGaussian(indices, values, sq, tl2, w, rTrain);
double currLoss = GLMOptimUtils.computeRidgeLoss(rTrain, w, lambda);
if (verbose) {
System.out.println("Iteration " + iter + ": " + " " + currLoss);
}
if (OptimUtils.isConverged(prevLoss, currLoss, epsilon)) {
break;
}
}
return GLMOptimUtils.getGLM(attrs, w, intercept, LinkFunction.IDENTITY);
}
