package jsat.classifiers.linear;
import java.util.List;
import jsat.classifiers.ClassificationDataSet;
import jsat.linear.DenseVector;
import jsat.linear.Vec;
import static java.lang.Math.*;
/**
* This class provides static helper methods that may be useful for various
* linear models.
*
* @author Edward Raff
*/
public class LinearTools
{
private LinearTools()
{
}
/**
* If the linear model performs logistic regression regularized by λ
* ||w||1, this method computes the smallest value of lambda that
* produces a weight vector of all zeros.
*
* Note, that the value returned depends on the data set size. If being used
* to initialize the value of λ for cross validation with k-folds,
* the value (k-1)/k * λ will be closer to the correct value of
* λ for each CV set.
*
* @param cds the data set that the model would be trained from
* @return the smallest value of λ that should produce all zeros.
*/
public static double maxLambdaLogisticL1(ClassificationDataSet cds)
{
/**
* This code was ripped out/modified from NewGLMNET. It follows the
* strategy laid out in Schmidt, M., Fung, G.,&Rosaless, R. (2009).
* Optimization Methods for L1-Regularization. Retrieved from
* http://www.cs.ubc.ca/cgi-bin/tr/2009/TR-2009-19.pdf , where we use
* the coordinate with the largest magnitude of the gradient
*/
/**
* if w=0, then D_part[i] = 0.5 for all i
*/
final double D_part_i = 0.5;
final int n = cds.getNumNumericalVars();
Vec delta_L = new DenseVector(n);
List X = cds.getDataVectors();
for (int i = 0; i < X.size(); i++)
{
double y_i = cds.getDataPointCategory(i) * 2 - 1;
Vec x = X.get(i);
delta_L.mutableAdd(D_part_i * y_i, x);
}
return max(abs(delta_L.max()), abs(delta_L.min())) / (cds.getSampleSize());
}
/**
* Many linear classifiers can be phrased in two equivalent forms, that only change the notation for the regularized. These forms are:
* C ∑i ℓ(w,xi) + Ω(w)
* and
* 1/N ∑i ℓ(w,xi) + λ Ω(w)
*
* This method converts the regularization parameter λ to the form used as C
*
* @param lambda the regularization parameter λ
* @param N the number of data points in the training set
* @return
*/
public static double lambda2C(double lambda, double N)
{
return 1/(lambda*N);
}
/**
* Many linear classifiers can be phrased in two equivalent forms, that only change the notation for the regularized. These forms are:
* C ∑i ℓ(w,xi) + Ω(w)
* and
* 1/N ∑i ℓ(w,xi) + λ Ω(w)
*
* This method converts the regularization parameter C to the form used as λ
*
* @param C the regularization parameter C
* @param N the number of data points in the training set
* @return
*/
public static double c2Lambda(double C, double N)
{
return 1/(C*N);
}
}
Ads help maintain this website.