weka[28]em源代码分析

《weka[28]em源代码分析》由会员上传分享，免费在线阅读，更多相关内容在应用文档-天天文库。

1、Weka[28]EM源代码分析作者：Koala++/屈伟EM算法在clusterers下面，提一下是因为我没有想到它竟然在这里，而且它的名字也太大了点，因为这里它只是与SimpleKMeans结合的算法。引自AndrewNg的LecturenotesmixturesofGaussiansandtheEMalgorithm：TheEM-algorithmisalsoreminiscentoftheK-meansclusteringalgorithm,exceptthatinsteadof“hard”clusterassignme

2、ntc(i),weinsteadhavethe“soft”assignmentw_j^(i).SimilartoK-means,itisalsosusceptibletolocaloptima,soreinitializingatseveraldifferentinitialparametersmaybeagoodidea。Soft指的是我们猜测是概率，取值在[0,1]区间，相反，“hard”猜测是指单个最好的猜测，可以取值在{0,1}或是{1,…,k}。英文原文：Theterm“soft”referstoourguesses

3、beingprobabilitiesandtakingvaluesin[0,1];incontrast,a“hard”guessisonethatrepresentsasinglebestguess(suchastakingvaluesin{0,1}or{1,…,k})下面的图来自NgAndrew和BishopChistopher，第一组图K-Means的猜测是两个点，而第二组图EM是对概率的猜测。另一点是刚才文中提到的，多个初始化点，在代码中也体现了。Ng在对EM算法收敛证明之后，解释如下：Hence,EMcausesthe

4、likelihoodtoconvergemonotonically.InourdescriptionoftheEMalgorithm,wesaidwe'drunituntilconvergence.Giventheresultthatwejustshowed,onereasonableconvergencetestwouldbetocheckiftheincreaseinl(theta)betweensuccessiveiterationsissmallerthansometoleranceparameter,andtodec

5、lareconvergenceifEMisimprovingl(theta)tooslowly.从buildCluster开始：if(data.checkForStringAttributes()){thrownewException("Can'thandlestringattributes!");}m_replaceMissing=newReplaceMissingValues();Instancesinstances=newInstances(data);instances.setClassIndex(-1);m_repl

6、aceMissing.setInputFormat(instances);data=weka.filters.Filter.useFilter(instances,m_replaceMissing);instances=null;m_theInstances=data;//calculateminandmaxvaluesforattributesm_minValues=newdouble[m_theInstances.numAttributes()];m_maxValues=newdouble[m_theInstances.n

7、umAttributes()];for(inti=0;i

8、最大值数组。privatevoidupdateMinMax(Instanceinstance){for(intj=0;j