package com.etsy.conjecture.model;
import com.etsy.conjecture.Utilities;
import com.etsy.conjecture.data.BinaryLabel;
import com.etsy.conjecture.data.LabeledInstance;
import com.etsy.conjecture.data.StringKeyedVector;
/**
* Hinge loss for binary classification tasks with y in {-1,1}.
* When threshold=1.0, one gets the loss used by SVM.
* When threshold=0.0, one gets the loss used by the Perceptron.
*/
public class Hinge extends UpdateableLinearModel<BinaryLabel> {
private static final long serialVersionUID = 1L;
private double threshold = 0.0;
public Hinge(SGDOptimizer optimizer) {
super(optimizer);
}
public Hinge(StringKeyedVector param, SGDOptimizer optimizer) {
super(param, optimizer);
}
@Override
public BinaryLabel predict(StringKeyedVector instance) {
double inner = param.dot(instance);
return new BinaryLabel(Utilities.logistic(inner));
}
@Override
public double loss(LabeledInstance<BinaryLabel> instance) {
double inner = param.dot(instance.getVector());
double label = instance.getLabel().getAsPlusMinus();
double z = inner * label;
if (z <= this.threshold) {
return this.threshold - z;
} else {
return 0.0;
}
}
@Override
public StringKeyedVector getGradients(LabeledInstance<BinaryLabel> instance) {
StringKeyedVector gradients = instance.getVector().copy();
double inner = param.dot(instance.getVector());
double label = instance.getLabel().getAsPlusMinus();
double z = inner * label;
if (z <= this.threshold) {
gradients.mul(-label);
return gradients;
} else {
return new StringKeyedVector();
}
}
@Override
protected String getModelType() {
return "hinge";
}
public Hinge setThreshold(double threshold) {
this.threshold = threshold;
return this;
}
}