[Home] DMelt member access. Java source code: jsat.classifiers.svm.extended.OnlineAMM
package jsat.classifiers.svm.extended;
import java.util.*;
import jsat.DataSet;
import jsat.classifiers.BaseUpdateableClassifier;
import jsat.classifiers.CategoricalData;
import jsat.classifiers.CategoricalResults;
import jsat.classifiers.DataPoint;
import jsat.distributions.Distribution;
import jsat.distributions.Gamma;
import jsat.distributions.LogUniform;
import jsat.exceptions.FailedToFitException;
import jsat.linear.DenseVector;
import jsat.linear.ScaledVector;
import jsat.linear.Vec;
import jsat.linear.VecWithNorm;
import jsat.parameters.Parameter;
import jsat.parameters.Parameterized;
import jsat.utils.DoubleList;
import jsat.utils.IndexTable;
import jsat.utils.IntList;
/**
* This is the Online variant of the Adaptive Multi-Hyperplane Machine (AMM)
* algorithm. It is related to linear SVMs where instead of having only a single
* weight vector, it is extended to multi-class problems by giving each class
* its own weight vector. It is further extended by allowing each class to
* dynamically add new weight vectors to increase the non-linearity of the
* solution.
* This algorithm works best for problems with a very large number of data
* points where traditional kernelized SVMs are prohibitively expensive to train
* due to computational cost.
*
* Unlike the batch variant, the online AMM algorithm has no convergence
* guarantees. However it still produces good results.
*
* See:
*
* - Wang, Z., Djuric, N., Crammer, K., & Vucetic, S. (2011). Trading
* representability for scalability Adaptive Multi-Hyperplane Machine for
* nonlinear Classification. In Proceedings of the 17th ACM SIGKDD
* international conference on Knowledge discovery and data mining - KDD ’11
* (p. 24). New York, New York, USA: ACM Press. doi:10.1145/2020408.2020420
* - Djuric, N., Lan, L., Vucetic, S., & Wang, Z. (2014). BudgetedSVM: A
* Toolbox for Scalable SVM Approximations. Journal of Machine Learning
* Research, 14, 3813–3817. Retrieved from
* here
*
*
* @author Edward Raff
*/
public class OnlineAMM extends BaseUpdateableClassifier implements Parameterized
{
private static final long serialVersionUID = 8291068484917637037L;
/*
* b/c of the batch learner we use a map, so that we dont have to think
* about how to handle re-assignment of data points to weight vectors.
* Also allows us to handle removed cases by checking if our owner is in the
* map. Use nextID to make sure we give every new vec a unique ID with
* respect to the class label
*/
protected List
Ads help maintain this website.