package jsat.classifiers.bayesian.graphicalmodel;
import java.util.HashMap;
import java.util.Map;
import java.util.Set;
import java.util.concurrent.ExecutorService;
import jsat.classifiers.CategoricalData;
import jsat.classifiers.CategoricalResults;
import jsat.classifiers.ClassificationDataSet;
import jsat.classifiers.Classifier;
import jsat.classifiers.DataPoint;
import jsat.classifiers.DataPointPair;
import jsat.classifiers.bayesian.ConditionalProbabilityTable;
import jsat.classifiers.bayesian.NaiveBayes;
import jsat.exceptions.FailedToFitException;
import jsat.utils.IntSet;
import static java.lang.Math.*;
/**
* A class for representing a Baysian Network (BN) for discrete variables. A BN use a graph to representing
* the relations between variables, and these links are called the structure. The structure of a BN must be
* specified by an expert using the {@link #depends(int, int) } method. The target class should be specified
* as the parent of the variables which have a causal relationship to it. These children of the target class
* should then have their own children specified. Once the structure has been specified, the network can be
* trained and used for classification.
* If the network structure has not been specified, or has no relationships for the target class, the BN will
* create an edge from the target class to every variable. If no edges were ever specified, this initialization
* of edges corresponds to a {@link NaiveBayes} implementation.
*
* @author Edward Raff
*/
public class DiscreteBayesNetwork implements Classifier
{
private static final long serialVersionUID = 2980734594356260141L;
/**
* The directed Graph that represents this BN
*/
protected DirectedGraph dag;
/**
* The Conditional probability tables for each variable
*/
protected Map cpts;
/**
* The class we are predicting
*/
protected CategoricalData predicting;
/**
* The prior probabilities of each class value
*/
protected double[] priors;
private boolean usePriors = DEFAULT_USE_PRIORS;
/**
* Whether or not the classifier should take into account the prior probabilities. Default value is {@value #DEFAULT_USE_PRIORS}.
*/
public static final boolean DEFAULT_USE_PRIORS = true;
public DiscreteBayesNetwork()
{
dag = new DirectedGraph();
}
public CategoricalResults classify(DataPoint data)
{
CategoricalResults cr = new CategoricalResults(predicting.getNumOfCategories());
int classId = data.numCategoricalValues();
//Use log proababilities to avoid underflow
double logPSum = 0;
double[] logProbs = new double[cr.size()];
for(int i = 0; i < cr.size(); i++)
{
DataPointPair dpp = new DataPointPair(data, i);
for(int classParent : dag.getChildren(classId))
logProbs[i] += log(cpts.get(classParent).query(classParent, dpp));
if(usePriors)
logProbs[i] += log(priors[i]);
logPSum += logProbs[i];
}
for(int i = 0; i < cr.size(); i++)
cr.setProb(i, exp(logProbs[i]-logPSum));
return cr;
}
/**
* Adds a dependency relation ship between two variables that will be in the network. The integer value corresponds
* the the index of the i'th categorical variable, where the class target's value is the number of categorical variables.
*
* @param parent the parent variable, which will be explained in part by the child
* @param child the child variable, which contributes to the conditional probability of the parent.
*/
public void depends(int parent, int child)
{
dag.addNode(child);
dag.addNode(parent);
dag.addEdge(parent, child);
}
public void train(ClassificationDataSet dataSet, boolean parallel)
{
train(dataSet);
}
public void train(ClassificationDataSet dataSet)
{
int classID = dataSet.getNumCategoricalVars();
if(classID == 0 )
throw new FailedToFitException("Network needs categorical attribtues to work");
predicting = dataSet.getPredicting();
priors = dataSet.getPriors();
cpts = new HashMap();
Set cptTrainSet = new IntSet();
if(dag.getNodes().isEmpty())
{
for(int i = 0; i < classID; i++)
depends(classID, i);
}
for(int classParent : dag.getChildren(classID))
{
Set depends = dag.getChildren(classParent);
ConditionalProbabilityTable cpt = new ConditionalProbabilityTable();
cptTrainSet.clear();
cptTrainSet.addAll(depends);
cptTrainSet.add(classParent);
cptTrainSet.add(classID);
cpt.trainC(dataSet, cptTrainSet);
cpts.put(classParent, cpt);
}
}
public boolean supportsWeightedData()
{
return false;
}
@Override
public Classifier clone()
{
throw new UnsupportedOperationException("Not supported yet.");
}
}
Ads help maintain this website.