資源描述:
《基于Ⅳ屬性選擇的隨機(jī)森林模型研究》由會(huì)員上傳分享,免費(fèi)在線閱讀,更多相關(guān)內(nèi)容在學(xué)術(shù)論文-天天文庫(kù)。
1、基于IV屬性選擇的隨機(jī)森林模型研究摘要隨著信息技術(shù)的迅速發(fā)展,眾多應(yīng)用領(lǐng)域如銀行金融業(yè)、電子商務(wù)、生物信息、網(wǎng)絡(luò)安全等產(chǎn)生了爆炸式的信息。不僅在數(shù)據(jù)規(guī)模上具有高維、海量的特征,在信息內(nèi)容上還具有冗余多、噪音多的特點(diǎn)。這樣的數(shù)據(jù)給挖掘技術(shù)帶來了巨大的挑戰(zhàn),尤其是處理數(shù)據(jù)流等問題時(shí),模型的實(shí)時(shí)性無法保障,使得更注重訓(xùn)練數(shù)據(jù)質(zhì)量的分類模型訓(xùn)練周期變長(zhǎng),精度下降。因此,如何有效的減小數(shù)據(jù)規(guī)模,提高數(shù)據(jù)質(zhì)量對(duì)提高分類模型的性能有著重要意義。本文針對(duì)屬性選擇及分類問題開展了以下工作的研究:(1)針對(duì)數(shù)據(jù)挖掘所面臨的挑戰(zhàn),分析了有效縮減數(shù)據(jù)規(guī)模是重要的可行方法,并在此基礎(chǔ)上概述了各類經(jīng)典屬性
2、選擇方法,探討了它們的特點(diǎn)與不足。(2)針對(duì)已有屬性選擇方法在處理高維、海量數(shù)據(jù)時(shí),時(shí)空性能與效果上的不足,分析了將WoE與IV指標(biāo)引入屬性選擇的可行性及存在的問題,在解決這些問題的基礎(chǔ)上提出了基于IV指標(biāo)的屬性選擇方法FS.IV,實(shí)驗(yàn)表明該算法是有效的,與經(jīng)典屬性選擇方法相比時(shí)空性能有明顯優(yōu)勢(shì),并具有一定的抗噪性。(3)針對(duì)屬性選擇后數(shù)據(jù)集出現(xiàn)的數(shù)據(jù)量大幅減少、優(yōu)勢(shì)屬性集中可能會(huì)導(dǎo)致的過擬合等問題,分析了解決手段,將FS.IV方法與隨機(jī)森林模型結(jié)合,提出了基于IV指標(biāo)的隨機(jī)森林模型,實(shí)驗(yàn)表明該模型與C4.5,樸素貝葉斯及經(jīng)過FS.IV約簡(jiǎn)的C4.5與樸素貝葉斯模型相比,在不損
3、失精度的情況下,時(shí)間性能大幅提升。(4)根據(jù)高維、海量、流數(shù)據(jù)等實(shí)際問題,對(duì)FS.IV及IV.RF模型做了適應(yīng)性改進(jìn),實(shí)驗(yàn)表明它們對(duì)高維、海量數(shù)據(jù)有著很好的處理效果。關(guān)鍵詞:屬性選擇,IV指標(biāo),隨機(jī)森林,數(shù)據(jù)挖掘TheResearchOilRandomForestBasedonIVFeatureSelectionAbstractWiththerapiddevelopmentofinformationtechnology,anexplosiveamountofdataisbroughtoutinthefieldslikebanking,financialservices,e-co
4、mmerce,bioinformaticsandnetworksecurity.Thesepracticaldatathatminingtasksfaceareoftenofhigh—dimension,redundantfeatures,aswellasnoises,whichmayleadtolowerprecisionandcostmoretime,especiallyinclassificationmodeling,sincehighqualitydataarepreferred.Thus,itwillbehelpfultousethosepredictivefeatu
5、resforimprovingtheperformancesInthisthesis,researchesarecarriedoutonfeatureselectionandclassificationasbelow:(1)Accordingtothechallengesthatdataminingfaces,apossiblewayistoreducehugedatasizeeffectivelysuchasfeatureselection.Wesummarizemostclassicalmethodsoffeatureselection,andpointouttheirch
6、aracteristicsaswellasweakpointsbasedontheanalysis.(2)Duetothedefectsoftraditionalmodelsthatwementioned,thefeasibilityandthedifficultyofusingWoEandIVasafeatureselectionmethodsareanalyzed.Undertheanalysis,afeatureselectionmodelFS-IVisproposedbasedontheIVindex.Experimentsshowthatthemodelperform
7、swithashortenedtimeandhassomenoiseimmunity.(3)Fortheproblemsthatfeatureselectionbrings,suchasthenotablecutondataandthegatheringofsuperiorfeatures,asuitableclassificationmodelIVoRFisproposed.ExperimentsshowthatthemodelhasasatisfiedtimecostWithlittle