package org.fnlp.ml.feature;
import gnu.trove.iterator.TIntFloatIterator;
import java.util.ArrayList;
import java.util.Arrays;
import org.fnlp.ml.classifier.bayes.Heap;
import org.fnlp.ml.classifier.bayes.ItemFrequency;
import org.fnlp.ml.types.sv.HashSparseVector;
/**
* 特征选择 保存选择后的特征
* @author sywu
*
*/
public class FeatureSelect {
private boolean isUseful[];
int size;
public FeatureSelect(int size){
this.size=size;
isUseful=new boolean[size];
Arrays.fill(isUseful, true);
}
/**
* 特征选择 卡方法 保留特征为每个类别前percent的最大卡方值特征的并集
* @param tf 词频类
* @param percent 保留百分比
*/
public void fS_CS(ItemFrequency tf,float percent){featureSelectionChiSquare(tf,percent);}
public void featureSelectionChiSquare(ItemFrequency tf,float percent){
int feaSize=tf.getFeatureSize();
int typeSize=tf.getTypeSize();
isUseful=new boolean[feaSize];
Arrays.fill(isUseful, false);
for(int j=0;j<typeSize;j++){
Heap<Integer> heap=new Heap<Integer>((int)(feaSize*percent),true);
for(int i=0;i<feaSize;i++){
double A,B,C,D,AC,AB,N;
N=tf.getTotal();
A=tf.getItemFrequency(i, j);
AB=tf.getFeatureFrequency(i);
AC=tf.getTypeFrequency(j);
B=AB-A;
C=AC-A;
D=N-AB-C;
double score=(A*D-B*C)*(A*D-B*C)/AB/(C+D);
heap.insert(score, i);
}
ArrayList<Integer> data=heap.getData();
for(int i=1;i<data.size();i++)
isUseful[data.get(i)]=true;
}
int total=0;
for(int i=0;i<feaSize;i++){
if(isUseful[i])
total++;
}
System.out.println("Feature Selection"+total+"/"+feaSize);
}
/**
* 特征选择 卡方法
* 对特征对每个类别求卡方，保留最大项，保留结果为前percent的特征
* @param tf 词频类
* @param percent 保留百分比
*/
public void fS_CS_Max(ItemFrequency tf,float percent){featureSelectionChiSquareMax(tf,percent);}
public void featureSelectionChiSquareMax(ItemFrequency tf,float percent){
int feaSize=tf.getFeatureSize();
int typeSize=tf.getTypeSize();
isUseful=new boolean[feaSize];
Arrays.fill(isUseful, false);
Heap<Integer> heap=new Heap<Integer>((int)(feaSize*percent),true);
for(int i=0;i<feaSize;i++){
double max=0;
for(int j=0;j<typeSize;j++){
double A,B,C,D,AC,AB,N;
N=tf.getTotal();
A=tf.getItemFrequency(i, j);
AB=tf.getFeatureFrequency(i);
AC=tf.getTypeFrequency(j);
B=AB-A;
C=AC-A;
D=N-AB-C;
double score=(A*D-B*C)*(A*D-B*C)/AB/(C+D)/AC/(B+D);
if(score>max)
max=score;
}
heap.insert(max, i);
}
ArrayList<Integer> data=heap.getData();
for(int i=1;i<data.size();i++)
isUseful[data.get(i)]=true;
int total=0;
for(int i=0;i<feaSize;i++){
if(isUseful[i])
total++;
}
System.out.println("Feature Selection"+total+"/"+feaSize);
}
/**
* 特征选择 信息增益
* @param tf 词频类
* @param percent 保留百分比
*/
public void fS_IG(ItemFrequency tf,float percent){featureSelectionInformationGain(tf,percent);}
public void featureSelectionInformationGain(ItemFrequency tf,float percent){
int feaSize=tf.getFeatureSize();
int typeSize=tf.getTypeSize();
isUseful=new boolean[feaSize];
Arrays.fill(isUseful, false);
Heap<Integer> heap=new Heap<Integer>((int)(feaSize*percent),true);
for(int i=0;i<feaSize;i++){
double ig=0;
for(int j=0;j<typeSize;j++){
double A,B,C,D,AC,AB,N;
N=tf.getTotal();
A=tf.getItemFrequency(i, j);
AB=tf.getFeatureFrequency(i);
AC=tf.getTypeFrequency(j);
B=AB-A;
C=AC-A;
D=N-AB-C;
ig+=-AC/N*Math.log(AC/N);
ig+=AB/N*A/N*Math.log(A/N);
ig+=(1-AB/N)*C/N*Math.log(C/N);;
}
heap.insert(ig, i);
}
ArrayList<Integer> data=heap.getData();
for(int i=1;i<data.size();i++)
isUseful[data.get(i)]=true;
int total=0;
for(int i=0;i<feaSize;i++){
if(isUseful[i])
total++;
}
System.out.println("Feature Selection"+total+"/"+feaSize);
}
public void noFeatureSelection(){
Arrays.fill(isUseful, true);
}
public HashSparseVector select(HashSparseVector vec){
HashSparseVector sv=new HashSparseVector();
TIntFloatIterator it=vec.data.iterator();
while(it.hasNext()){
it.advance();
if(isUseful[it.key()])
sv.put(it.key(), it.value());
}
return sv;
}
}