基于PU学习算法虚假评论识别研究优秀参考文档.pdfVIP

计算 机研究 与发展 DOI：10．7544／issn1000—1239．2015．2O131473 JournalofComputerResearchandDevelopment 52(3)：639—648，2015 基于 PU学习算法的虚假评论识别研究 任亚峰 姬东鸿 张红斌 哥 兰 (武汉大学计算机学院 武汉 430072) (renyafeng@whu．edu．cn) DeceptiveReviewsDetection Basedon PositiveandUnlabeledLearning RenYafeng，JiDonghong，ZhangHongbin，andYinLan (SchoolofComputer，WuhanUniversity，Wuhan430072) Abstract Identifyingdeceptivereviewshasimportanttheoreticalmeaningandpracticalvalue．W hile previousworksfocuson someheuristicrulesortraditionalsupervisedmethods．Recentresearchhas shownthathumans cannotdirectly identify deceptive reviewsby their prior knowledge． Human— annotateddatasetmustcontain somemislabeled examples．Due tothe difficultyofhuman labeling needed for supervised learning，the problem remainsto behighly challenging． There are some ambiguousreviews(wecallthem spyexamples)，whichareeasilymislabeled．Thekeyofidentifying deceptivereview ishow todealwith thesespyreviews．Basedon sometruthfulreviewsandalarge amountofunlabeledreviews，anovelapproach，calledmixingpopulation andindividualnaturePU learning，isproposed． Firstly，somereliablenegative examplesare identified from the unlabeled dataset．Secondly，some representative positive examplesand negative examples are generated by integratinglatentdirichletallocationandK—means．Thirdly，allspyexamplesareclusteredintomany groupsbasedondirichletprocessmixturemodel，andtwo schemes(populationnatureandindividual nature)aremixedtodeterminethecategorylabelofspyexamples．Finally，multiplekernellearningis presentedtobuildthefina1classifier．Experimentalresultsdemonstratethatourproposedmethods can effectivelyidentifydeceptivereviews，andoutperform thecurrentbaselines． Keywords deceptivereviews；supervisedlearning；positiveandunlabeled (PU)learning；Dirichlet processmixturemodel(DPM M)；multiplekernellearning (M KL) 摘 要 识别虚假评论有着重要的理论意义与现实价值．先前工作集中于启发式策略和传统的全