半监督聚类算法对于流和多密度数据.pdf

《半监督聚类算法对于流和多密度数据.pdf》由会员上传分享，免费在线阅读，更多相关内容在学术论文-天天文库。

1、北京理工大学Semi-SupervisedClusteringAlgorithmsforStreamingandMulti-DensityDataWalidSaidAbdelhamidAtwa2015年6月Semi-SupervisedClusteringAlgorithmsforStreamingandMulti-DensityDataCandidateName:WalidSaidAbdelhamidAtwaSchoolorDepartment:SchoolofComputerScienceSuperv

2、isor:ProfessorDr.LiKanChair,ThesisCommittee:Prof.XuyanTuDegreeApplied:DoctorofPhilosophyMajor:ComputerScienceandTechnologyDegreeby:BeijingInstituteofTechnologyTheDateofDefence:June–2015中图分类号：TP311.1UDC分类号：004.62半监督聚类算法对于流和多密度数据作者姓名阿特瓦学院名称计算机学院指导教师李侃教授答辩委员

3、会主席涂序彦教授申请学位工学博士学科专业计算机科学与技术学位授予单位北京理工大学论文答辩日期2015年6月研究成果声明本人郑重声明：所提交的学位论文是我本人在指导教师的指导下进行的研究工作获得的研究成果。尽我所知，文中除特别标注和致谢的地方外，学位论文中不包含其他人已经发表或撰写过的研究成果，也不包含为获得北京理工大学或其它教育机构的学位或证书所使用过的材料。与我一同工作的合作者对此研究工作所做的任何贡献均已在学位论文中作了明确的说明并表示了谢意。特此申明。签名：日期：北京理工大学博士学位论文Abstra

4、ctClusteringisoneofthemostcommondataminingtasks,usedfrequentlyfordatacategorizationandanalysisinbothindustryandacademia.Inmanydomainswhereclusteringisapplied,somepriorknowledgeisavailableeitherintheformoflabeleddata(specifyingthecategorytowhichaninstanceb

5、elongs)orpairwiseconstraintsonsomeoftheinstances(specifyingwhethertwoinstancesshouldbeinsameordifferentclusters).Thefocusofourresearchisonsemi-supervisedclustering,wherewestudyhowpriorknowledgecanbeincorporatedintoclusteringalgorithms.Semi-supervisedclust

6、eringaimstoimprovetheclusteringperformancebyconsideringusersupervisionintheformofpairwiseconstraints.However,mostcurrentalgorithmsarepassiveinthesensethatpairwiseconstraintsareprovidedbeforehandandselectedrandomly.Thismayleadtotheuseofconstraintsthatarere

7、dundant,unnecessary,orevenharmfultotheclusteringresults.Forthosereasons,wewouldliketooptimizetheselectionoftheconstraintsforsemi-supervisedclustering.Moreover,semi-supervisedclusteringalgorithmsimposesseveralchallengestobeaddressed,suchasdealingwithmulti-

8、densitydata,howtohandletheevolvingpatternsthatareimportantcharacteristicsofstreamingdatawithdynamicdistributions,capableofperformingfastandincrementalprocessingofdataobjects,andsuitablyaddressingtimeandmemorylimitat