协同过滤外文文献翻译

1、外文:IntroductiontoRecommenderSystemApproachesofCollaborativeFiltering:NearestNeighborhoodandMatrixFactorization“Weareleavingtheageofinformationandenteringtheageofrecommendation.”Likemanymachinelearningtechniques,arecommendersystemmakespredictionbasedonusers'historicalbehaviors.Specifically,it'

2、;stopredictuserpreferenceforasetofitemsbasedonpastexperience.Tobuildarecommendersystem,themosttwopopularapproachesareContent-basedandCollaborativeFiltering.Content-basedapproachrequiresagoodamountofinformationofitems'ownfeatures,ratherthanusingusers'interactionsandfeedbacks.Forexample,itcanb

3、emovieattributessuchasgenre,year,director,actoretc.,ortextualcontentofarticlesthatcanextractedbyapplyingNaturalLanguageProcessing.CollaborativeFiltering,ontheotherhand,doesn'tneedanythingelseexceptusershistoricalpreferenceonasetofitems.Becauseit'sbasedonhistoricaldata,thecoreassumptionhereis

4、thattheuserswhohaveagreedinthepasttendtoalsoagreeinthefuture.Intermsofuserpreference,itusuallyexpressedbytwocategories.ExplicitRating,isarategivenbyausertoanitemonaslidingscale,like5starsforTitanic.Thisisthemostdirectfeedbackfromuserstoshowhowmuchtheylikeanitem.ImplicitRating,suggestsuserspreference

5、indirectly,suchaspageviews,clicks,purchaserecords,whetherornotlistentoamusictrack,andsoon.Inthisarticle,Iwilltakeacloselookatcollaborativefilteringthatisatraditionalandpowerfultoolforrecommendersystems.NearestNeighborhoodThestandardmethodofCollaborativeFilteringisknownasNearestNeighborhoodalgorithm.

6、ThereareuserbasedCFanditem-basedCF.Let'sfirstlooUater-basedCF.Wehaveannmmatrixofratings,withuseru?i=1,.nanditemp?,j=1,-m.Nowwewanttopredicttheratingr?iftargetuserididnotwatch/rateanitemj.Theprocessistocalculatethesimilaritiesbetweentargetuseriandallotherusers,selectthetopXsimilarusers,andtakethe

7、weightedaverageofratingsfromtheseXuserswithsimilaritiesasweights.£SimilariesiUj,_“numberofratingsWhiledifferentpeoplemayhavedifferentbaselineswhengivingratings,somepeopletendtogivehighscoresgenerally,someareprettystricteventhoughtheyaresatisfiedwithitems.Toavoidthisbias,wecansubtracteachuser

8、9;saverageratingofatemswhencomputingweightedaverage,andadditbackfortargetuser,shownasbelow.£Similanes（uh心）（r灯一'）3=0+numberofratingsTwowaystocalculatesimilarityarePearsonCorrelationandCosineSimilarity.2（%一衣）（%一X）PearsonCorrelation:Sim/<,=JjE（"产一（均一一产CosineSimilarity:翼办=Basically,thei

9、deaistofindthemostsimilaruserstoyourtargetuser（nearestneighbors）andweighttheirratingsofanitemasthepredictionoftheratingofthisitemfortargetuser.Withoutknowinganythingaboutitemsandusersthemselves,wethinktwousersaresimilarwhentheygivethesameitemsimilarratings.Analogously,forItembasedCF,wesaytwoitemsare

10、similarwhentheyreceivedsimilarratingsfromasameuser.Then,wewillmakepredictionforatargetuseronanitembycalculatingweightedaverageofratingsonmostXsimilaritemsfromthisuser.OnekeyadvantageofItem-basedCFisthestabilitywhichisthattheratingsonagivenitemwillnotchangesignificantlyovertime,unlikethetastesofhuman

11、beings.Therearequiteafewlimitationsofthismethod.Itdoesnhandlesparsitywellwhennooneintheneighborhoodtargetuser.Also,itratedanitemthatiswhatyouaretryingtopredictforsnotcomputationalefficientasthegrowthofthenumberofusersandproducts.MatrixFactorizationSincesparsityandscalabilityarethetwobiggestchallenge

12、sforstandardCFmethod,itcomesamoreadvancedmethodthatdecomposetheoriginalsparsematrixtolowdimensionalmatriceswithlatentfactors/featuresandlesssparsity.ThatisMatrixFactorization.Besidesolvingtheissuesofsparsityandscalability,there'sanintuitiveexplanationofwhyweneedlow-dimensionalmatricestorepresent

13、users'preference.AusergavegoodratingstomovieAvatar,Gravity,andInception.Theyarenotnecessarily3separateopinionsbutshowingthatthisusersmightbeinfavorofSci-FimoviesandtheremaybemanymoreSci-Fimoviesthatthisuserwouldlike.Unlikespecificmovies,latentfeaturesisexpressedbyhigher-levelattributes,andSci-Fi

14、categoryisoneoflatentfeaturesinthiscase.Whatmatrixfactorizationeventuallygivesusishowmuchauserisalignedwithasetoflatentfeatures,andhowmuchamoviefitsintothissetoflatentfeatures.Theadvantageofitoverstandardnearestneighborhoodisthateventhoughtwousershaven'tratedanysamemovies,it'sstillpossibleto

15、findthesimilaritybetweenthemiftheysharethesimilarunderlyingtastes,againlatentfeatures.Toseehowamatrixbeingfactorized,firstthingtounderstandisSingularValueDecomposition(SVD)BasedonLinearAlgebra,anyrealmatrixRcanbedecomposedinto3matricesU,2,andV.Continuingusingmovieexample,Uisannruser-latentfeaturemat

16、rix,Visanmr>movie-latentfeaturematrix.2isanrxrdiagonalmatrixcontainingthesingularvaluesoforiginalmatrix,simplyrepresentinghowimportantaspecificfeatureistopredictuserpreference.R=UZVTUWM,EGIRrxVeIRrxmTosortthevaluesof2bydecreasingabsolutevalueandtruncatematrix2tofirstkdimensions(ksingularvalues),w

17、ecanreconstructthematrixasmatrixA.TheselectionofkshouldmakesurethatAisabletocapturethemostofvariancewithintheoriginalmatrixR,sothatAistheapproximationofR,A弋R.ThedifferencebetweenAandRistheerrorthatisexpectedtobeminimized.ThisisexactlythethoughtofPrincipleComponentAnalysis.WhenmatrixRisdense,UandVcou

18、ldbeeasilyfactorizedanalytically.However,amatrixofmovieratingsissupersparse.Althoughtherearesomeimputationmethodstofillinmissingvalues,wewillturntoaprogrammingapproachtojustlivewiththosemissingvaluesandfindfactormatricesUandV.InsteadoffactorizingRviaSVD,wearetryingfindUandVdirectlywiththegoalthatwhe

19、nUandVmultipliedbacktogethertheoutputmatrixR'istheclosestapproximationofRandnomoreasparsematrix.ThisnumericalapproximationisusuallyachievedwithNon-NegativeMatrixFactorizationforrecommendersystemssincethereisnonegativevaluesinratings.Seetheformulabelow.Lookingatthepredictedratingforspecificuseran

20、ditem,itemiisnotedasavectorq?anduseruisnotedasavectorp?suchthatthedotproductofthesetwovectorsisthepredictedratingforuseruonitemi.ThisvalueispresentedinthematrixR'atrowuandcolumni.PredictedRatings:r：。=RHowdowefindoptimalq?andp?Likemostofmachinelearningtask,alossfunctionisdefinedtominimizethecosto

21、ferrors.min一浦+小瓦II?+帅iP)MJr?isthetrueratingsfromoriginaluser-itemmatrix.OptimizationprocessistofindtheoptimalmatrixPcomposedbyvectorp?andmatrixQcomposedbyvectorq?inordertominimizethesumsquareerrorbetweenpredictedratingsr?andthetrueratingsr?Also,L2regularizationhasbeenaddedtopreventoverfittingofusera

22、nditemvectors.It'salsoquitecommontoaddbiastermwhichusuallyhas3majorcomponents:averageratingofallitems以,averageratingofitemiminus?),以(notedasbaverageratinggivenbyuseruminusu(notedasb?.min口同上1f也2（入一工人+*+几+加）+"帆俨+右+层+如）OptimizationAfewoptimizationalgorithmshavebeenpopulartosolveNon-NegativeFac

23、torization.AlternativeLeastSquareisoneofthem.Sincethelossfunctionisnon-convexinthiscase,there'snowaytoreachaglobalminimum,whileitstillcanreachagreatapproximationbyfindinglocalminimums.AlternativeLeastSquareistoholduserfactormatrixconstant,adjustitemfactormatrixbytakingderivativesoflossfunctionan

24、dsettingitequalto0,andthensetitemfactormatrixconstantwhileadjustinguserfactormatrix.Repeattheprocessbyswitchingandadjustingmatricesbackandforthuntilconvergence.IfyouapplyScikit-learnNMFmodel,youwillseeALSisthedefaultsolvertouse,whichisalsocalledCoordinateDescent.Pysparkalsooffersprettyneatdecomposit

25、ionpackagesthatprovidesmoretuningflexibilityofALSitself.SomeThoughtsCollaborativeFilteringprovidesstrongpredictivepowerforrecommendersystems,andrequirestheleastinformationatthesametime.However,ithasafewlimitationsinsomeparticularsituations.First,theunderlyingtastesexpressedbylatentfeaturesareactuall

26、ynotinterpretablebecausethereisnocontent-relatedpropertiesofmetadata.Formovieexample,itdoesn'tnecessarilytobegenrelikeSci-Fiinmyexample.Itcanbehowmotivationalthesoundtrackis,howgoodtheplotis,andsoon.CollaborativeFilteringislackoftransparencyandexplainabilityofthislevelofinformation.Ontheotherhan

27、d,CollaborativeFilteringisfacedwithcoldstart.Whenanewitemcomingin,untilithastoberatedbysubstantialnumberofusers,themodelisnotabletomakeanypersonalizedrecommendations.Similarly,foritemsfromthetailthatdidn'tgettoomuchdata,themodeltendstogivelessweightonthemandhavepopularitybiasbyrecommendingmorepo

28、pularitems.It'susuallyagoodideatohaveensemblealgorithmstobuildamorecomprehensivemachinelearningmodelsuchascombiningcontent-basedfilteringbyaddingsomedimensionsofkeywordsthatareexplainable,butweshouldalwaysconsiderthetradeoffbetweenmodel/computationalcomplexityandtheeffectivenessofperformanceimpr

29、ovement.中文翻译推荐系统介绍协同过滤的方法：最近邻域和矩阵分解我们正在离开信息时代，而进入推荐时代。”像许多机器学习技术一样，推荐系统根据用户的历史行为进行预测。具体来说，是根据过去的经验来预测用户对一组商品的偏好。要构建推荐系统，最流行的两种方法是基于内容的过滤和协作过滤。基于内容的方法需要大量项目自身功能的信息，而不是使用用户的交互和反馈。例如，它可以是电影属性（例如流派，年份，导演，演员等）或可以通过应用自然语言处理提取的文章的文本内容。另一方面，协作过滤除了用户对一组项目的历史偏好之外，不需要任何其他操作。因为它是基于历史数据的，所以这里的核心假设是，过去已经同意的用户将来也会

30、倾向于也同意。就用户偏好而言，它通常由两类表示。明确评分，是用户按滑动比例对某项商品的价格，例如泰坦尼克号的评分为5星。这是用户最直接的反馈，表明他们对商品的喜爱程度。隐含评价，间接建议用户偏好，例如页面浏览量，点击次数，购买记录，是否收听音乐曲目等等。在本文中，我将仔细研究协作过滤，它是推荐系统的传统且功能强大的工具。最近的邻居协作过滤的标准方法称为最近邻算法”。有基于用户的CF和基于项目的CF。让我们先来看看基于用户的CF。我们有一个nXm的评分矩阵，用户u?,i=1,.n,项目p?,j=1,.m。现在，如果目标用户i没有对项目j进行观看/评分，我们现在要预测评分r?o该过程将计算目标用户

31、i与所有其他用户之间的相似度，选择排名靠前的X个相似用户，并将来自这X个具有相似性的用户的评分的加权平均值作为权重。£Si>nilaries（uit也），灯k为二Zrnumberofratings尽管不同的人给由评分时可能会有不同的基准，但是有些人通常会给由高分，有些人即使对项目感到满意也很严格。为了避免这种偏差，我们可以在计算加权平均值时减去每个用户对所有项目的平均评分，然后将其加回到目标用户身上，如下所示。WSmilanes（Ui，唠（陶-4）=?-+'numberofratings一计算相似度的两种方法是皮尔森相关和余弦相似度。£（%一门）1%3Pear

32、sonCorrelation:5加（,妣）='./£（为一一S物一门：mErijrkf,.r：-r；>=lCosineSimilarity:Sim（%,uQ-=-IfJIftlfmmJ斗遇4VJ=ij=i基本上，该想法是找到与您的目标用户（最接近的邻居）最相似的用户，并权衡他们对某项的评价，以此作为对目标用户的评价。在不了解商品和用户本身的情况下，我们认为两个用户在给同一个商品相似的评分时是相似的。类似地，对于基于项目的CF,我们说两个项目在收到来自同一用户的相似评分时是相似的。然后，我们将通过计算来自该用户的大多数X个类似商品的评分的加权平均值，来预测该商品的目标用户

33、。基于项目的CF的一个关键优势是稳定性，即与人类的口味不同，给定项目的评级不会随着时间的推移而发生显着变化。此方法有很多限制。当附近没有人对您要为目标用户预测的商品进行评分时，它不能很好地处理稀疏性。而且，随着用户和产品数量的增长，它的计算效率也不高。矩阵分解由于稀疏性和可伸缩性是标准CF方法的两个最大挑战，因此由现了一种更高级的方法，该方法将原始稀疏矩阵分解为具有潜在因子/特征且稀疏性较低的低维矩阵。那就是矩阵分解。除了解决稀疏性和可伸缩性问题之外，还有一个直观的解释，说明为什么我们需要低维矩阵来表示用户的偏好。用户对电影阿凡达，重力和盗梦空间给予了很高的评价。它们不一定是3个独立的意见，而

34、是表明该用户可能更喜欢科幻电影，并且该用户可能想要更多的科幻电影。与特定电影不同，潜在功能由更高级别的属性表示，在这种情况下，科幻类别是潜在功能之一。矩阵分解最终给我们的是用户与一组潜在特征对齐的程度，以及一部电影在这组潜在特征中的适应程度。与标准最近邻区相比，它的优势在于，即使两个用户未对任何一部电影进行评级，但如果他们共享相似的基本口味(又是潜在特征)，仍然有可能找到它们之间的相似性。要了解矩阵如何分解，首先要了解的是奇异值分解(SVD)o基于线性代数，可以将任何实矩阵R分解为3个矩阵U,2和Vo以电影示例为例，U是nxr用户潜伏特征矩阵，V是命r电影潜伏特征矩阵。2是一个rM对角矩阵，包含原始矩阵的奇异值，仅表示特定功能对预测用户偏好的重要性。R=LFLVtSeIRryVeIRrxm为了通过减少绝对值对2的值进行排序并将矩阵2截断为前k个维（k个奇异值）,我们可以将矩阵重构为矩阵Ao选择k应该确保A能够捕获最大的方差在原始矩阵R内，A是R的近似值，A=R。A和R之间的差是期望最小化的误差。这正是主成分分析的思想。当矩阵R是致密的时，U和V可以很容易地解析分解。但是，电影分级矩阵超级稀疏。尽管存在一些填补缺失值的插补方法，但我们将转向一种编