基于集成學習的不平衡數(shù)據(jù)分類的研究及應用.pdf

1、大連理工大學學位論文獨創(chuàng)性聲明作者鄭重聲明：所呈交的學位論文，是本人在導師的指導下進行研究工作所取得的成果。盡我所知，除文中已經注明引用內容和致謝的地方外，本論文不包含其他個人或集體已經發(fā)表的研究成果，也不包含其他已申請學位或其他用途使用過的成果。與我一同工作的同志對本研究所做的貢獻均已在論文中做了明確的說明并表示了謝意。若有不實之處，本人愿意承擔相關法律責任a學位論文題目：廛五紐蔓芝z壺至塑紐盤絲霉：美查塑幽作者簽名：一團孚笈一一日期

2、：叢!竺年—厶月上二日ResearchandApplicationofImbalancedDataClassificationAlgorithmsBasedonEnsembleLearningAbstractInrecentyears，imbalanceddataclassificationhasbecomeoneoftheimportantresearchissuesindataminingImbalanceddatarefersto

3、thevariousclassesofdatasetsdisparitiesinthenumberofsamplesinwhichtheclassthathaslesssamplesiScalledminorityclasstheclassthathasmoresamplesiscalledmajorityclassThetraditionalclassificationalgorithmswhendealingwithbalanced

4、ataclassificationproblemscanoftenplayaneffectiveperformanceadvantage，however，whendealingwithimbalanceddataclassification，correctclassificationrateofminorityclasssamplesareoftenlessthantheclassificationaccuracyofthemajori

5、tyclasssamplesInmanypracticalapplications，suchasfrauddiagnostics，faultdetection，textdetection，spamfiltering，etc，areusuallymoreconcemedabouttheaccuracyoftheminorityclasssampleclassificationTherefore，thestudyofhowtoimprove

6、theclassificationperformanceofimbalanceddataandthegeneralizationabilityofclassifierhastheimportantvalueandpracticalsignificanceInordertoimprovetheperformanceofimbalanceddataclassification，peoplehasmadealotofimprovementon

7、thebasisofthetraditionalclassificationalgorithmsTheseimprovementsaremainlyconcentratedintwoaspects：dataplaneandalgorithmsTheimprovementofdatalevelismainlythatresamplesthesamplesofdatasetsuchasrandomoversampling，SMOTEalgo

8、rithm，andonesidesamplingalgorithmDatasetsbecomebasicbalancebetweendifferentclassesbychangingitsdistributioncharacteristics，andthenitclassedbythetraditionalclassificationalgorithmsAlgorithmlevel，whilemaintainingthedistrib

9、utionoftheoriginaldatamodifytheinternalstructureofthealgorithmSOthatitcanadapttounbalanceddataclassification，suchastheclassificationalgorithmbasedoncost—sensitive，improvedSVM，ensemblelearningalgorithmIntheimprovedalgorit

10、hm，theensemblelearningcanexhibitgoodclassificationperformanceands打onggeneralizationabilityItcanfurtherimprovetheclassificationperformanceintegratedlearningbyimprovingthesingleclassifierandcoordinatingthediversitybetweent

11、hebaseclassifiersBasedontheaboveconsideration，thispaperhascarriedonthefollowingseveralaspectswork：first，inthedatalevel，analyzesandsummarizestheonesidesamplingandSMOTEalgorithm，aimingattheexistingproblemofalgorithm，propos