[免费论文]探讨微博用户转发预测的的特证
取[第五章]微博用户转发特征选择实验[第六章]Filter特征子集与Wrapper特征子集对比[第七章]微博转发预测的的结论与参考文献摘要微博用户转发预测探究对社交网络的信息传播和推荐探究具有重要的学术价值,在公共舆论管理.个性化推荐.市场营销等方面具有重要的应用价值.本文主要探究微博中普通用户转发一条微博的重要影响因素,目的是更好的理解用户的转发行为,同时提高转发预测效果.本文通过剖析微博用户转发行为的影响因素,综合相关文献,汇总了影响用户转发行为的众多特征,在规模较大的真实微博数据集上实现了特征提取与特征选择,建立因子分解机预测模型,在测试集上对用户的转发行为做出预测,通过对比探究了各种特征和特征组合在微博用户转发行为预测上的有效性.本文的工作主要有四个方面:(1)综合相关文献,剖析汇总了大量影响用户转发行为的特征,并在真实的新浪微博数据集上实现了特征提取.(2)通过每次使用一组类型特征训练建立因子分解机预测模型的方式,探究了用户特征.作者特征.微博特征.兴趣特征和社交特征五种不同类型特征对模型预测性能的影响.实验表明,兴趣特征和微博特征对模型预测性能的影响最大.(3)对微博特征全集实现了Filter特征选择和Wrapper特征选择,探究了各种特征和特征子集对模型预测性能的影响.实验表明,转发相似度特征与分类预测的相关性最高.Wrapper方法选择的最优特征子集,在几乎保证预测效果的同时,大幅降低了特征维度,提高了运行效率.(4)利用预测性能最好的最优特征子集建立因子分解机预测模型,对用户的转发行为进行预测,预测精度达到了89.0%,F1度量达到了66.8%,AUC面积达到了95.0%.关键字:转发预测,特征提取,特征选择,因子分解机ABSTRACTWeibousers'forwardingpredictionresearchhasimportantacademicvaluefortheresearchofsocialnetwork'sinformationdisseminationandrecommendation.Ithasimportantapplicationvalueinpublicopinionmanagement,personalizedrecommendationandmarketing,etc.Thispapermainlystudiestheimportantinfluencingfactorsofordinaryusers'forwardingbehaviorinWeibo.Thepurposeistobetterunderstandtheforwardingbehaviorofusersandimprovetheeffectofforwardingprediction.ThispaperanalyzestheinfluencingfactorsofWeibousers'forwardingbehavior,synthesizesrelatedliteratures,andsummarizesmanyfeaturesthataffectusers'forwardingbehavior.Thenonthelarge-scalerealWeibodatasets,featureextractionandfeatureselectionareperformed,aFactorizationMachinespredictionmodelisestablished,andtheforwardingbehaviorofusersinthetestsetispredicted.Basedonthepredictionresults,thispapercomparestheeffectivenessofvariousfeaturesandfeaturecombinationsinWeibouser'sforwardingbehaviorprediction.Theworkofthispaperhasfourmainaspects:1.Basedonthecomprehensiveliteratures,itanalyzesandsummarizesalargenumberoffeaturesaffectingusers'forwardingbehavior,andimplementsfeatureextractionontherealSinaWeibodataset.2.ThroughusingasetoftypefeaturestrainingtoestablishaFactorizationMachinespredictionmodel,itstudiestheimpactonusers'forwardingbehavioroffivedifferenttypesoffeatureswhichincludeuserfeatures,authorfeatures,Weibofeatures,interestfeaturesandsocialfeatures.ExperimentsshowthatinterestfeaturesandWeibofeatureshavethegreatestimpactonmodel'spredictionperformance.3.FilterfeatureselectionandWrapperfeatureselectionareimplementedonthecompletesetofWeibofeatures.Theeffectsofvariousfeaturesandfeaturesubsetsonthepredictionperformanceofthemodelarestudied.Experimentsshowthattheforwardingsimilarityfeaturehasthehighestcorrelationwiththeclassificationprediction.TheoptimalfeaturesubsetselectedbytheWrappermethodgreatlyreducesthefeaturedimensionandenhancestheoperationalefficiencywhilealmostguaranteeingthepredictioneffect.4.UsingtheFactorizationMachinespredictionmodelestablishedbytheoptimalfeaturesubsetwiththebestpredictionperformancetopredicttheuser'sforwardingbehavior,thepredictionaccuracyreaches89.0%,theF1metricreaches66.8%,andtheAUCareareaches95.0%.Keywords:Forwardingprediction,Featureextraction,Featureselection,FactorizationMachines目录摘要ABSTRACT目录第一章绪论1.1探究时代和意义1.1.1探究时代1.1.2探究意义1.2国内外探究现状1.3主要探究内容1.4本文组织结构第二章相关知识介绍2.1特征提取2.2特征选择2.3因子分解机2.4本章小结第三章数据预处理3.1数据集说明3.2数据预处理3.2.1数据清理13.2.2负样本识别3.2.3文本分词3.3本章小结第四章特征剖析与提取4.1影响因素剖析4.2特征提取4.3特征表示4.4不同类型特征预测性能对比4.5本章小结第五章特征选择实验5.1实验设置5.1.1实验环境5.1.2评价指标5.2Filter特征选择5.2.1卡方检验方法5.2.2实验与结果剖析5.3Wrapper特征选择5.3.1递归特征消除方法5.3.2实验与结果剖析5.4本章小结第六章Filter特征子集与Wrapper特征子集对比6.1模型预测性能对比6.2特征对比6.3本章小结总结与展望参考文献致谢
原文链接:http://www.jxszl.com/lwqt/mflw/16605.html