mutikdfu.java

Java语言实现的支持向量机的源码。对研究人工智能的朋友有用。

package core;

import java.text.NumberFormat;

import java.io.BufferedReader;
import java.io.FileReader;
import java.io.FileWriter;

import util.MyMath;
import weka.core.Attribute;
import weka.core.FastVector;
import weka.core.Instance;
import weka.core.Instances;
import weka.core.matrix.EigenvalueDecomposition;
import weka.core.matrix.Matrix;

import weka.classifiers.Classifier;
import util.MyClassifier;

public class MutiKDFU {
	//训练集
	private Instances srcData;
	//核矩阵M,采用RBF核,用来表示映射到高维空间F后的类间距
	public Matrix kernelM;
	private Matrix sumK;
	//每类类别的样本数目,这里处理多类样本
	private double num[];
	
	
	
	//核矩阵的特征向量
	private Matrix eigVec;
	//核矩阵的特征值
	private Matrix eigVal;
	////////////////////////
	//alpha特征向量
	private Matrix alpha;
	//参数omit,我们选取为两类之间的最小距离?还是最大? (先取最大)
	private int rankKM;
	private Matrix oness ;
	private double omit;
	private double[] der;
	
	//在构造函数中确定rankKM的值，不使用自适应的方法
	public MutiKDFU(Instances data,double[] inder,int inputMRank, double getomit){
		data.sort(data.numAttributes()-1);
		srcData = new Instances(data);
		num = new double[data.numClasses()];
		kernelM = new Matrix(data.numInstances(),data.numInstances());
		oness = new Matrix(data.numInstances(),data.numInstances());
		der = inder;
		rankKM = inputMRank; //
		omit=getomit;
	}
	//计算各类的样本数目 ok
	private void calSpNum(){
		int i;
		for(i=0;i<srcData.numClasses();i++)
		{
			num[i] = 0;			
		}
		for(i = 0;i<srcData.numInstances();i++)
		{
			num[(int)srcData.instance(i).classValue()]++;
		
		}			
		
		
	}
	//建立核矩阵M.
	private void buildKM()
	{		
		sumK = new Matrix(srcData.numInstances(),srcData.numInstances());
		Matrix ones = new Matrix(srcData.numInstances(),srcData.numInstances()); //均值矩阵
		omit = Double.MIN_VALUE; //omit为核函数的参数.迩衡是用启发式来确定这个值.
		MyMath mh = new MyMath();//调用迩衡的mymath类,运用里面的数学方法
		//先算出omit的值,用于下面的核变换
	
		for(int i=0;i<srcData.numInstances();i++)
		{
			for(int j=0;j<srcData.numInstances();j++)
			{
				double v = mh.CalSim(srcData.instance(i), srcData.instance(j));
				if(srcData.instance(i).classValue()!=srcData.instance(j).classValue())
				{
					omit = Math.max(v,omit);
				}
			}
		}
		//omit=omit*0.5;
		//System.out.println("mutiKDFU omit is "+omit);
		//omit = 10;
		//这个ALLMi是用来保存每个Mi向量.有多少类就有多少个Mi向量
		
		Matrix ALLMi = new Matrix(srcData.numInstances(),num.length);
		Matrix M1 = new Matrix(srcData.numInstances(),1);
		Matrix M2 = new Matrix(srcData.numInstances(),1);
		
		//Mi可以指代M1,M2,M3...Mk
		Matrix Mi = new Matrix(srcData.numInstances(),1);
		//这个n维向量Mc是用来记录均值向量的
		Matrix Mc = new Matrix(srcData.numInstances(),1);
		//用来暂时存放向量的
		Matrix tempM = new Matrix(srcData.numInstances(),1);
		
		
		//for (int nn=0;nn<der.length;nn++)
		//	System.out.println("the der[i] in BulidKM is "+der[nn]);
		//build Mi~Mk &&ALLMi
		//************************
		//这里要对照一下那个公式,是不是有新版本的公式
		int numcont = 0;
		for(int n = 0; n < num.length; n++)
		{
			double value = 0;  //每次value要清零.
			for(int i = 0; i < srcData.numInstances();i++)
			{
				for(int j =0; j < num[n];j++)
				{
					value = value + mh.CalEachDemSim(srcData.instance(i), srcData.instance(j+numcont), der, omit);
					
				}
				
				value =(1.0/(double)num[n])*value;
				ALLMi.set(i, n, value);
			}
			numcont+=(int)num[n];
		}
		//计算均值向量Mc
		
		for(int i =0; i < num.length; i++)
		{
			
			tempM = ALLMi.getMatrix(0,srcData.numInstances()-1, i, i);
			Mc.plusEquals(tempM);
		}
		Mc.timesEquals(1.0/(double)num.length);  // 1/k*sumMc
		
		//计算核矩阵M
		for(int i = 0; i < num.length; i++)
		{
			tempM = ALLMi.getMatrix(0, srcData.numInstances()-1,i,i);
			Matrix M = tempM.minus(Mc);  //Mi-Mc
			Matrix Mt = M.transpose();   //(Mi-Mc)T
			Matrix MMt=M.times(Mt);
			kernelM.plusEquals(MMt);
			
		}
		kernelM.timesEquals(1.0/(double)num.length);
					
		//对kernelM矩阵做规整化,确保kernelM的秩为满秩
		Matrix IN = new Matrix(kernelM.getRowDimension(),kernelM.getColumnDimension());
		//build IN		
		for(int i=0;i<IN.getColumnDimension();i++)
			//IN.set(i, i, 0.00000000000001);
			IN.set(i, i, 0.000000001);
			
		kernelM=kernelM.plus(IN);
		
		System.out.println("******* the rank of MutiKDFU is "+kernelM.rank());
		
	}	
	private void calEig(){
		EigenvalueDecomposition ED = kernelM.eig();
		eigVal = ED.getD();
		eigVec = ED.getV();
		new MyMath().quickSort(eigVal, eigVec);		//特征值和特征向量排一下序
		//**************************************************************************
		/*
		//rankKM  = 0;                                //过滤特征值过小的特征向量
		double minV =eigVal.get(0, 0)/100;      //设定特征值的阀值为(最大特征值/1000).但除最大特征值,其他都没有一个比阀值大...其他的特征向量趋于0
		for(int i =0; i < eigVal.getRowDimension();i++) //根据阀值来选取比较大的特征向量和特征值
		{
			if(eigVal.get(i, i)<minV)break;
			rankKM++;
		}
		rankKM = Math.min(rankKM, srcData.numInstances());	
	    //下面这句可以去掉,暴力设定取110维.正常的是应该根据上面程序段来确定rankKM的值.
        // rankKM =150;
		*///************************************************************************
		System.out.println("rankKM for MutiKDFU is "+rankKM);
    }
	//压缩原来的核矩阵来构造新的核矩阵
	
	//构造alpha特征向量
	private void calAlpha(){
		alpha = new Matrix(srcData.numInstances(),rankKM);
		alpha = eigVec.getMatrix(0, eigVec.getRowDimension()-1,0,rankKM-1);		
	}
	
	//映射新样本集
	public Instances deInstances(Instances newData){
		MyMath mh = new MyMath();
		//System.out.println("the omit for kdfu is "+omit);
		//当前样本集和学习样本集的核矩阵
		Matrix newKernel = new Matrix(newData.numInstances(),srcData.numInstances());
		for(int i = 0; i < newData.numInstances(); i++)
		{
			double sum = 0.0;
			for(int j=0;j<srcData.numInstances();j++)
			{
				//double val = mh.CalEachDemSim(newData.instance(i), srcData.instance(j),der,omit);
				
				double val = Math.exp(-mh.CalSim(newData.instance(i), srcData.instance(j))/omit);
				newKernel.set(i, j, val);
				sum += val;
			}
			for(int j=0;j<srcData.numInstances();j++) //迩衡说这个是使得矩阵均值为0
			{
				newKernel.set(i, j, newKernel.get(i, j)-sum/srcData.numInstances());
			}			
		}
		Matrix newSum = sumK.getMatrix(0, 0, 0, sumK.getRowDimension()-1);
		double sumV = 0.0;
		for(int i =0; i <sumK.getRowDimension();i++) sumV += newSum.get(0, i);
		for(int i =0; i <sumK.getRowDimension();i++) newSum.set(0, i, newSum.get(0,i)-sumV/srcData.numInstances());
		/*for(int i =0; i <newData.numInstances();i++)
		{
			for(int j=0; j<srcData.numInstances();j++)
			{
				newKernel.set(i, j, newKernel.get(i, j)-newSum.get(0, j));  //也是矩阵均值化.
			}
		}*/
		newKernel = newKernel.times(alpha);  //新样本集映射到F空间
		
		FastVector fv = new FastVector();
		//设置特征属性
		for(int i=0;i<newKernel.getColumnDimension();i++)
		{
			fv.addElement(new Attribute("F"+i));			
		}
		//设置类属性
		FastVector classAttribute = new FastVector();
		
		for(int i =0; i<newData.numClasses();i++)
		{
			classAttribute.addElement(""+i);
		}
		fv.addElement(new Attribute("class",classAttribute));
		//创建实例集
		Instances disData = new Instances("RD_"+srcData.relationName(),fv,newData.numInstances());
		for(int i =0;i<newKernel.getRowDimension();i++)
		{
			Instance f = new Instance(newKernel.getColumnDimension()+1);
			for(int j=0;j<newKernel.getColumnDimension();j++)
			{
				f.setValue(j, newKernel.get(i, j));
			}
			f.setValue(newKernel.getColumnDimension(), (int)newData.instance(i).classValue());
			disData.add(f);
		}
		disData.setClassIndex(newKernel.getColumnDimension());
		return disData;
	}
	
	
	//执行
	public void excute()throws Exception{
		
		calSpNum();    //先计算各类样本数目,用于下面建M矩阵等计算		
		buildKM();     //建立M矩阵
		calEig();      //计算M矩阵的特征向量和特征值
		calAlpha();     //算出alpha

	}
			
}