主动学习:
主动学习的过程:需要分类器与标记专家进行交互。一个典型的过程:
(1)基于少量已标记样本构建模型
(2)从未标记样本中选出信息量最大的样本,交给专家进行标记
(3)将这些样本与之前样本进行融合,并构建模型
(4)重复执行步骤(2)和步骤(3),直到stopping criterion(不存在未标记样本或其他条件)满足为止
模拟思路:
1. 将数据分为label 和 unlabel数据集
2. 将 unlabel 分为100个一组,每组样本数组分别求出熵值,按照熵值排序,取前5个样本,添加到 label样本之中
package demo; import java.io.FileReader;
import java.util.ArrayList;
import java.util.Collections;
import java.util.Random;
import weka.classifiers.Evaluation;
import weka.classifiers.bayes.NaiveBayes;
import weka.core.Instance;
import weka.core.Instances;
import weka.core.converters.ConverterUtils.DataSource; //将测试用例,按照熵值进行排序
class InstanceSort implements Comparable<InstanceSort>{
public Instance instance;
public double entropy; public InstanceSort( Instance instance, double entropy){
this.instance = instance;
this.entropy = entropy;
}
@Override
public int compareTo(InstanceSort o) {
// TODO Auto-generated method stub
if (this.entropy < o.entropy){
return 1;
}else if ( this.entropy > o.entropy){
return -1;
} return 0;
}
} public class ActiveLearning { public static Instances getInstances( String fileName) throws Exception{
Instances data = new Instances (new FileReader(fileName));
data.setClassIndex(data.numAttributes()-1);
return data;
} //计算熵
public static double computeEntropy(double predictValue){
double entropy = 0.0;
if ( 1-predictValue < 0.000000001d || predictValue < 0.000000001d){
return 0;
}else {
return -predictValue*(Math.log(predictValue)/Math.log(2.0d))-(1-predictValue)*(Math.log(1-predictValue)/Math.log(2.0d));
}
} public static void classify(Instances train, Instances test) throws Exception{
NaiveBayes classifier = new NaiveBayes();
//训练模型
classifier.buildClassifier(train); //评价模型
Evaluation eval = new Evaluation(test);
eval.evaluateModel(classifier, test);
System.out.println(eval.toClassDetailsString());
} //不确定采样
public static Instances uncertaintySample(Instances labeled, Instances unlabeled, int start, int end) throws Exception{
//用有标签的先训练模型
NaiveBayes classifier = new NaiveBayes();
classifier.buildClassifier(labeled);
//按照熵进行排序
ArrayList <InstanceSort> l = new ArrayList<InstanceSort>(); for (int i = start; i < end; i++) {
double result = classifier.classifyInstance(unlabeled.instance(i));
double entropy = computeEntropy (result);
InstanceSort is = new InstanceSort(unlabeled.instance(i), entropy);
l.add(is);
}
//按照熵值进行排序
Collections.sort(l); DataSource source = new DataSource("NASA//pc1.arff");
Instances A = source.getDataSet();
Instances chosenInstances = new Instances(A, 0);
//每100个里面选择5个熵值最小的实例
for(int i = 0; i < 5; i++){
chosenInstances.add(l.get(i).instance);
} return chosenInstances;
} //采样
public static void sample( Instances instances, Instances test) throws Exception{
Random rand = new Random(1023);
instances.randomize(rand);
instances.stratify(10);
Instances unlabeled = instances.trainCV(10, 0);
Instances labeled = instances.testCV(10, 0); int iterations = unlabeled.numInstances() / 100 +1; for ( int i=0; i< iterations-1 ; i++){
//每100个里面选择5个熵值最小的实例
//100个一组
Instances resultInstances = uncertaintySample(labeled, unlabeled, i*100, (i+1)*100);
for (int j = 0; j < resultInstances.numInstances(); j++){
labeled.add(resultInstances.instance(j));
}
classify(labeled, test);
} Instances resultInstances = uncertaintySample(labeled, unlabeled, (iterations-1)*100, unlabeled.numInstances()); for (int j = 0; j < resultInstances.numInstances(); j++){
labeled.add(resultInstances.instance(j));
} classify(labeled, test); } public static void main(String[] args) throws Exception{
// TODO Auto-generated method stub
Instances instances = getInstances("NASA//pc1.arff"); //10-fold cross validation
Random rand = new Random(1023);
instances.randomize(rand);
instances.stratify(10);
Instances train = instances.trainCV(10, 0);
Instances test = instances.testCV(10, 0);
// System.out.println(train.numInstances());
// System.out.println(test.numInstances()); sample(train,test); } }