資源描述:
《基于特征子集區(qū)分度與支持向量機的特征選擇算法-論文.pdf》由會員上傳分享,免費在線閱讀,更多相關(guān)內(nèi)容在行業(yè)資料-天天文庫。
1、第37卷第8期計算機學(xué)報Vo1.37No.82014年8月CHINESEJOURNAL0FCOMPUTERSAug.2014基于特征子集區(qū)分度與支持向量機的特征選擇算法謝娟英”謝維信’”(陜西師范大學(xué)計算機科學(xué)學(xué)院西安710062)(深圳大學(xué)信息工程學(xué)院ATR國家重點實驗室廣東深圳518060)摘要考慮特征之間的相關(guān)性對于其類間區(qū)分能力的影響,提出了一種新的特征子集區(qū)分度衡量準(zhǔn)則——DFS(DiscernibilityofFeatureSubsets)準(zhǔn)則.該準(zhǔn)則考慮特征之間的相關(guān)性,通過計算特征子集中全部特征對于分類的聯(lián)合貢獻來判斷特征子集的類間
2、辨別能力大小,不再只考慮單個特征對于分類的貢獻.結(jié)合順序前向、順序后向、順序前向浮動和順序后向浮動4種特征搜索策略,以支持向量機(SupportVectorMachines,SVM)為分類工具,引導(dǎo)特征選擇過程,得到4種基于DFS與SVM的特征選擇算法.其中在順序前/后向浮動搜索策略中,首先根據(jù)DFS準(zhǔn)則加入/去掉特征到特征子集中,然后在浮動階段根據(jù)所得臨時SVM分類器的分類性能決定剛加入/去掉特征的去留.UCI機器學(xué)習(xí)數(shù)據(jù)庫數(shù)據(jù)集的對比實驗測試表明,提出的DFS準(zhǔn)則是一種很好的特征子集類間區(qū)分能力度量準(zhǔn)則;基于DFS與SVM的特征選擇算法實現(xiàn)了有
3、效的特征選擇;與其他同類算法相比,基于DFS準(zhǔn)則與SVM的特征選擇算法具有非常好的泛化性能,但其所選特征子集的規(guī)模不一定是最好的.關(guān)鍵詞特征選擇;支持向量機;相關(guān)性;特征子集區(qū)分度;特征區(qū)分度中圖法分類號TP18DOI號10.3724/SP.J.1016.2014.01704SeveralFeatureSelectionAlgorithmsBasedontheDiscernibilityofaFeatureSubsetandSupportVectorMachinesXIEJuan-YingXIEWei-Xin。’”(SchoolofComputer
4、Science,ShaanxiNormalUniversity,Xi’an710062)2(SchoolofInformationEngineering,NationalLaboratoryofATR,ShenzhenUniversity,Shenzhen,Guangdong518060)AbstractToconsidertheinfluenceofthecorrelationbetweenfeaturesontheirdiscernibilitybetweenclasses,anewcriterionwasproposedinthispaper
5、toevaluatethediscernibilityofafeaturesubset.WereferredtothiscriterionasDFSfortheshortofthediscernibilityoffeaturesubsets.DFSconsidersthecorrelationbetweenfeaturesbycomputingthediscernibilityofthewholefeatureSUbsetbetweenclasses,SOthatitcanmeasurethecontributionofthewholefeatur
6、esubsettotheclassificationnotonlythatofonefeature.FourfeatureselectionalgorithmswereputforwardbycombiningtheDFS,respectively,withthesequentialforwardsearch,sequentialbackwardsearch,sequentialforwardfloatingsearch,andthesequentialbackwardfloatingsearchstrategieswheresupportvect
7、ormachines(SVM)wereusedasaclassificationtooltoguidethefeatureselectionprocedure,especiallyinthesequentialforward/backwardfloatingsearchprocedureswhereafeaturewasfirstaddedto/deletedfromthefeaturesubsetusingtheDFScriterion,thenitwasdeletedfrom/calledbackduringthefloatingprocedu
8、redependingontheaccuracyofthecorrespondingtemporarySVMclassif