Figureb2b程序
b2b程序 时间:2021-04-13 阅读:()
AHybridB2BAppRecommenderSystemAlexandruOprea1,ThomasHornung2,Cai-NicolasZiegler3,HolgerEggs1,andGeorgLausen21SAPCommercialPlatform,St.
Leon-Rot&SAPResearch,Darmstadt,Germany{alexandru.
dorin.
oprea,holger.
eggs}@sap.
com2InstituteofComputerScience,Albert-Ludwigs-Universit¨atFreiburg,Germany{hornungt,lausen}@informatik.
uni-freiburg.
de3AmericanExpress,PAYBACKGmbH,M¨unchen,Germanycai-nicolas.
ziegler@payback.
netAbstract.
RecommendersystemsareintegraltoB2Ce-commerce,withlittleusesofarinB2B.
WepresentaliverecommendersystemthatoperatesinadomainwhereusersarecompaniesandtheproductsbeingrecommendedB2Bapps.
Besidesoperatinginanentirenewdomain,theSAPStorerecommenderisbasedonaweightedhybriddesign,makinguseofanovelcondence-basedweightingschemeforcombiningratings.
Evaluationshaveshownthatoursystemperformssignicantlybetterthanatop-sellerrecommenderbenchmark.
1IntroductionandMotivationTheSAPStorecaterstoSMEcompaniesthataimtodrivetheirbusinessviaB2Bapps,e.
g.
,forcustomerrelationmanagementorcompliance.
Manyoftheseappsaregearedtowardsspecicindustriesandtheirneeds.
Asthenumberofpartnersproducingthemisgrowing,soisthenumberofappsinthestoreitselfandthusthecomplexityfortheuser(whorepresentsacompany)toactuallyndwhatheislookingfor.
Toactivelyhelptheuser,weproposeahybridrecommendersystemthataddressesexactlytheneedsofthisspecicB2Bscenario.
Thesystemputstousebothknowledge-based,collaborative,andcontent-basedsub-recommenders.
Moreover,wepresentanovelhybridweightingscheme[1]thatincorporatescon-dencescoringforthepredictionsproduced,sothatsub-recommenderscontributeforrecommendationsaccordingtotheircondenceweight.
Thesystemisliveandcanbeusedbylogged-inusers1.
Wehaveconductedempiricalevaluationsviahold-outtestingthatshowthattherecommenderout-performsthenon-personalizedtop-sellerrecommender.
2RecommenderSystemArchitectureThearchitectureoftherecommenderisdepictedinFigure1.
Overall,wehavethreedierentinformationsourcesforgeneratingnewrecommendations:the1Seehttp://store.
sap.
comF.
Daniel,P.
Dolog,andQ.
Li(Eds.
):ICWE2013,LNCS7977,pp.
490–493,2013.
cSpringer-VerlagBerlinHeidelberg2013AHybridB2BAppRecommenderSystem491Knowledge-basedFilter(KBF)UserProfilesAppProfilesTRXDataUser-ItemCFItem-ItemCFContent-basedAugmentationContent-basedAugmentationItem-ItemMatrixUser-UserMatrixWeightedMeanRecommendationList12a2b34Fig.
1.
SAPStorerecommendersystemarchitectureuserproles(e.
g.
,companysize,industry,country),theappproles(e.
g,sup-portedindustries,businessareas),andthetransactionalcustomerdata(e.
g.
,salesorders,downloads).
Initially,theknowledge-basedcomponentltersthelistofrelevantappsbyasetofplausibilityrulesresultinginanunsortedsetofcandidateapps(1).
Thesearefedtoanitem-item(2a)anduser-itemcollaborativelter(CF),see(2b)[2].
Todealwiththecold-startproblemincaseswhereonlysparseratingsareavailableforapps,acontent-basedaugmentationschemecomputessimilaritiesbasedonthecosinesimilaritymeasure[3]betweenpropertiesoftheapps.
Forusersthatarenewtothesystem,thesimilaritycanbedeterminedbycomparingtheirprolestootherusersbasedontheircosinesimilarity.
Thisway,thetwomatriceswillcontainmeaningfulentriesforallusersandappsknowntothesystem,andrecommendationsgetmorepersonalizedoncemorecontextdataisavailable.
ThescoresofthetwoCFalgorithmsarecombinedbyaweightedmean(cf.
Section2.
1),andasortedtop-krecommendationlistisreturned.
Thecalculationofthematricesisdoneo-lineasthecomputationisquadraticinthenumberofusersorapps,respectively.
2.
1WeightingbyCondenceScoresThescoreofarecommendedappisbasedonaweightedmeanoftheconstituentitem-itemanduser-itemscores.
Eachofthesegivesanestimateofhowmuchausermightlikeanapp;e.
g.
,Eq.
1showshowapredictionscorefortheitem-itemcaseisdeterminedforappamforuseru:Theratingsru(b)ofuforappsb∈Ru492A.
Opreaetal.
hehasalreadyratedareweightedbytheirsimilaritytoam,denoteds(b,am),asanindicatorifthisappmightberelevantfortheuser2.
pi(u,am)=b∈Rus(am,b)·ru(b)b∈Rus(am,b)(1)Now,foreachrecommenderscoreacondencescoreiscalculated,denotedciandcurespectively,whichisbasedonthenumberofsupportingitemsorusersofeachprediction.
Theseweightsareusedtodeterminetheoverallscorep:p(u,am)=ci·pi(u,am)+cu·pu(u,am)ci+cu(2)Thecondencescorecuforthepredictionpu(u,am)tellsushowreliableapre-dictionis.
Itgrowswithagrowingnumberofsupportingdatapoints:Foreachuserui,wecalculatethez-scoreofhissimilaritywithourcurrentuseru.
Wenowsumthesez-scoresimilaritiesforallkusersinuseru'sneighborhood[2].
Thesumisdividedbykandtheresultingvaluegivesustheaveragenormal-izedsimilarityofalltheuserswhoseratingshavebeentakenintoaccountforpu(u,am).
Thesameisdonefortheitem-basedcase.
Sincewearemakinguseofstandardz-scores,thelinearcombinationshowninEq.
2basedonthetwocondenceweightsissound.
Thecondenceschemerepre-sentsapowerfulmeanstoadjustthehybridrecommender'sweightingaccordingtothepredictedreliabilityofeachofthetwosub-recommenders.
3PerformanceEvaluationInordertotesttheperformanceofthepresentedhybridrecommenderusingournovelcondence-basedweightingscheme,weconductedanempiricalevaluationwithreal-worlddataof5,233users(e.
g.
,companiesregisteredforandusingtheSAPStore)havingpurchasedorexpressedinterestin615appsolutions.
ThefrequencydistributioninFig.
2(a)showsleadsperapp,i.
e.
,howmanycompanieshavepurchasedorexpressedinterestineachapp,sortedindescendingorder.
Thelog-logplottedgraphexhibitsapower-lawdistribution,soasmallnumberofappsattractsahighnumberofleads.
ThisisconrmedbyFig.
2(b),showingthatthetop-5appsaccumulate29%ofallleads,andtop-100capture90%.
Wethusconjecturethatanon-personalizedtop-sellerrecommender,whichonlyrecommendsthetop-Nmostpopularapps,willperformverywell.
Weadoptedahold-outcross-validationapproachfortesting,whereoneratingrvofauseriswithheldandallothersareusedtodenehisproleandcalcu-latepredictions,aimingtorecommendexactlyrv.
Forbaselining,wecomparedourrecommender'sperformancewiththatofthetop-sellerrecommender.
Theevaluationtaskforeachofthetworecommenderswastoproducealistoftop-Nrecommendationsandcountinhowmanycasestheproducedlistcontainedrv.
TheevaluationisshowninTab.
1.
Allresultsexhibitstatisticalsignicanceatthepη(a)йййййййййййη(b)Fig.
2.
Log-logfrequencydistributionofleadsperapp(a)andcumulativeshareofleadsbynumberofapps(b)Table1.
PerformancebenchmarkresultsTop-1Top-3Top-5Top-10Hybridrecommender10.
9%24.
4%33.
5%51.
2%Top-seller6.
6%18.
9%27.
6%43.
4%4ConclusionandOutlookWehavepresentedourrecommenderforthenewdomainofB2Bapps,makinguseofanovelhybridweightedschemebasedoncondencescoring.
OurrstevaluationshaveshownverypromisingresultsandthesystemhasgoneliveintooperationaluseatSAP.
Inthefuture,wewanttotunetherecommendingalgorithmsfurtherandaimatdoingthematrixcalculationsinreal-time,usingHANA[4],SAP'snewhigh-performancein-memorydatabase.
References1.
Burke,R.
:HybridWebRecommenderSystems.
In:Brusilovsky,P.
,Kobsa,A.
,Nejdl,W.
(eds.
)AdaptiveWeb2007.
LNCS,vol.
4321,pp.
377–408.
Springer,Heidelberg(2007)2.
Adomavicius,G.
,Tuzhilin,A.
:TowardtheNextGenerationofRecommenderSys-tems:ASurveyoftheState-of-the-ArtandPossibleExtensions.
IEEETrans.
Knowl.
DataEng.
17(6),734–749(2005)3.
Baeza-Yates,R.
A.
,Ribeiro-Neto,B.
A.
:ModernInformationRetrieval-TheCon-ceptsandTechnologyBehindSearch,2ndedn.
PearsonEducationLtd.
,Harlow(2011)4.
F¨arber,F.
,May,N.
,Lehner,W.
,Groe,P.
,M¨uller,I.
,Rauhe,H.
,Dees,J.
:TheSAPHANADatabase–AnArchitectureOverview.
IEEEDataEng.
Bull.
35(1),28–33(2012)
Leon-Rot&SAPResearch,Darmstadt,Germany{alexandru.
dorin.
oprea,holger.
eggs}@sap.
com2InstituteofComputerScience,Albert-Ludwigs-Universit¨atFreiburg,Germany{hornungt,lausen}@informatik.
uni-freiburg.
de3AmericanExpress,PAYBACKGmbH,M¨unchen,Germanycai-nicolas.
ziegler@payback.
netAbstract.
RecommendersystemsareintegraltoB2Ce-commerce,withlittleusesofarinB2B.
WepresentaliverecommendersystemthatoperatesinadomainwhereusersarecompaniesandtheproductsbeingrecommendedB2Bapps.
Besidesoperatinginanentirenewdomain,theSAPStorerecommenderisbasedonaweightedhybriddesign,makinguseofanovelcondence-basedweightingschemeforcombiningratings.
Evaluationshaveshownthatoursystemperformssignicantlybetterthanatop-sellerrecommenderbenchmark.
1IntroductionandMotivationTheSAPStorecaterstoSMEcompaniesthataimtodrivetheirbusinessviaB2Bapps,e.
g.
,forcustomerrelationmanagementorcompliance.
Manyoftheseappsaregearedtowardsspecicindustriesandtheirneeds.
Asthenumberofpartnersproducingthemisgrowing,soisthenumberofappsinthestoreitselfandthusthecomplexityfortheuser(whorepresentsacompany)toactuallyndwhatheislookingfor.
Toactivelyhelptheuser,weproposeahybridrecommendersystemthataddressesexactlytheneedsofthisspecicB2Bscenario.
Thesystemputstousebothknowledge-based,collaborative,andcontent-basedsub-recommenders.
Moreover,wepresentanovelhybridweightingscheme[1]thatincorporatescon-dencescoringforthepredictionsproduced,sothatsub-recommenderscontributeforrecommendationsaccordingtotheircondenceweight.
Thesystemisliveandcanbeusedbylogged-inusers1.
Wehaveconductedempiricalevaluationsviahold-outtestingthatshowthattherecommenderout-performsthenon-personalizedtop-sellerrecommender.
2RecommenderSystemArchitectureThearchitectureoftherecommenderisdepictedinFigure1.
Overall,wehavethreedierentinformationsourcesforgeneratingnewrecommendations:the1Seehttp://store.
sap.
comF.
Daniel,P.
Dolog,andQ.
Li(Eds.
):ICWE2013,LNCS7977,pp.
490–493,2013.
cSpringer-VerlagBerlinHeidelberg2013AHybridB2BAppRecommenderSystem491Knowledge-basedFilter(KBF)UserProfilesAppProfilesTRXDataUser-ItemCFItem-ItemCFContent-basedAugmentationContent-basedAugmentationItem-ItemMatrixUser-UserMatrixWeightedMeanRecommendationList12a2b34Fig.
1.
SAPStorerecommendersystemarchitectureuserproles(e.
g.
,companysize,industry,country),theappproles(e.
g,sup-portedindustries,businessareas),andthetransactionalcustomerdata(e.
g.
,salesorders,downloads).
Initially,theknowledge-basedcomponentltersthelistofrelevantappsbyasetofplausibilityrulesresultinginanunsortedsetofcandidateapps(1).
Thesearefedtoanitem-item(2a)anduser-itemcollaborativelter(CF),see(2b)[2].
Todealwiththecold-startproblemincaseswhereonlysparseratingsareavailableforapps,acontent-basedaugmentationschemecomputessimilaritiesbasedonthecosinesimilaritymeasure[3]betweenpropertiesoftheapps.
Forusersthatarenewtothesystem,thesimilaritycanbedeterminedbycomparingtheirprolestootherusersbasedontheircosinesimilarity.
Thisway,thetwomatriceswillcontainmeaningfulentriesforallusersandappsknowntothesystem,andrecommendationsgetmorepersonalizedoncemorecontextdataisavailable.
ThescoresofthetwoCFalgorithmsarecombinedbyaweightedmean(cf.
Section2.
1),andasortedtop-krecommendationlistisreturned.
Thecalculationofthematricesisdoneo-lineasthecomputationisquadraticinthenumberofusersorapps,respectively.
2.
1WeightingbyCondenceScoresThescoreofarecommendedappisbasedonaweightedmeanoftheconstituentitem-itemanduser-itemscores.
Eachofthesegivesanestimateofhowmuchausermightlikeanapp;e.
g.
,Eq.
1showshowapredictionscorefortheitem-itemcaseisdeterminedforappamforuseru:Theratingsru(b)ofuforappsb∈Ru492A.
Opreaetal.
hehasalreadyratedareweightedbytheirsimilaritytoam,denoteds(b,am),asanindicatorifthisappmightberelevantfortheuser2.
pi(u,am)=b∈Rus(am,b)·ru(b)b∈Rus(am,b)(1)Now,foreachrecommenderscoreacondencescoreiscalculated,denotedciandcurespectively,whichisbasedonthenumberofsupportingitemsorusersofeachprediction.
Theseweightsareusedtodeterminetheoverallscorep:p(u,am)=ci·pi(u,am)+cu·pu(u,am)ci+cu(2)Thecondencescorecuforthepredictionpu(u,am)tellsushowreliableapre-dictionis.
Itgrowswithagrowingnumberofsupportingdatapoints:Foreachuserui,wecalculatethez-scoreofhissimilaritywithourcurrentuseru.
Wenowsumthesez-scoresimilaritiesforallkusersinuseru'sneighborhood[2].
Thesumisdividedbykandtheresultingvaluegivesustheaveragenormal-izedsimilarityofalltheuserswhoseratingshavebeentakenintoaccountforpu(u,am).
Thesameisdonefortheitem-basedcase.
Sincewearemakinguseofstandardz-scores,thelinearcombinationshowninEq.
2basedonthetwocondenceweightsissound.
Thecondenceschemerepre-sentsapowerfulmeanstoadjustthehybridrecommender'sweightingaccordingtothepredictedreliabilityofeachofthetwosub-recommenders.
3PerformanceEvaluationInordertotesttheperformanceofthepresentedhybridrecommenderusingournovelcondence-basedweightingscheme,weconductedanempiricalevaluationwithreal-worlddataof5,233users(e.
g.
,companiesregisteredforandusingtheSAPStore)havingpurchasedorexpressedinterestin615appsolutions.
ThefrequencydistributioninFig.
2(a)showsleadsperapp,i.
e.
,howmanycompanieshavepurchasedorexpressedinterestineachapp,sortedindescendingorder.
Thelog-logplottedgraphexhibitsapower-lawdistribution,soasmallnumberofappsattractsahighnumberofleads.
ThisisconrmedbyFig.
2(b),showingthatthetop-5appsaccumulate29%ofallleads,andtop-100capture90%.
Wethusconjecturethatanon-personalizedtop-sellerrecommender,whichonlyrecommendsthetop-Nmostpopularapps,willperformverywell.
Weadoptedahold-outcross-validationapproachfortesting,whereoneratingrvofauseriswithheldandallothersareusedtodenehisproleandcalcu-latepredictions,aimingtorecommendexactlyrv.
Forbaselining,wecomparedourrecommender'sperformancewiththatofthetop-sellerrecommender.
Theevaluationtaskforeachofthetworecommenderswastoproducealistoftop-Nrecommendationsandcountinhowmanycasestheproducedlistcontainedrv.
TheevaluationisshowninTab.
1.
Allresultsexhibitstatisticalsignicanceatthepη(a)йййййййййййη(b)Fig.
2.
Log-logfrequencydistributionofleadsperapp(a)andcumulativeshareofleadsbynumberofapps(b)Table1.
PerformancebenchmarkresultsTop-1Top-3Top-5Top-10Hybridrecommender10.
9%24.
4%33.
5%51.
2%Top-seller6.
6%18.
9%27.
6%43.
4%4ConclusionandOutlookWehavepresentedourrecommenderforthenewdomainofB2Bapps,makinguseofanovelhybridweightedschemebasedoncondencescoring.
OurrstevaluationshaveshownverypromisingresultsandthesystemhasgoneliveintooperationaluseatSAP.
Inthefuture,wewanttotunetherecommendingalgorithmsfurtherandaimatdoingthematrixcalculationsinreal-time,usingHANA[4],SAP'snewhigh-performancein-memorydatabase.
References1.
Burke,R.
:HybridWebRecommenderSystems.
In:Brusilovsky,P.
,Kobsa,A.
,Nejdl,W.
(eds.
)AdaptiveWeb2007.
LNCS,vol.
4321,pp.
377–408.
Springer,Heidelberg(2007)2.
Adomavicius,G.
,Tuzhilin,A.
:TowardtheNextGenerationofRecommenderSys-tems:ASurveyoftheState-of-the-ArtandPossibleExtensions.
IEEETrans.
Knowl.
DataEng.
17(6),734–749(2005)3.
Baeza-Yates,R.
A.
,Ribeiro-Neto,B.
A.
:ModernInformationRetrieval-TheCon-ceptsandTechnologyBehindSearch,2ndedn.
PearsonEducationLtd.
,Harlow(2011)4.
F¨arber,F.
,May,N.
,Lehner,W.
,Groe,P.
,M¨uller,I.
,Rauhe,H.
,Dees,J.
:TheSAPHANADatabase–AnArchitectureOverview.
IEEEDataEng.
Bull.
35(1),28–33(2012)
亚洲云-浙江高防BGP，至强铂金8270，提供自助防火墙管理，超大内存满足你各种需求
官方网站:点击访问亚洲云官网618活动方案：618特价活动（6.18-6.30）全站首月活动月底结束！地区：浙江高防BGPCPU：至强铂金8270主频7 默频3.61 睿频4.0核心：8核（最高支持64核）内存：8G（最高支持128G）DDR4 3200硬盘：40G系统盘+80G数据盘带宽：上行：20Mbps/下行：1000Mbps防御：100G（可加至300G）防火墙：提供自助 天机盾+金盾 管...
pia云低至20/月，七折美国服务器
Pia云是一家2018的开办的国人商家，原名叫哔哔云，目前整合到了魔方云平台上，商家主要销售VPS服务，采用KVM虚拟架构 ，机房有美国洛杉矶、中国香港和深圳地区，洛杉矶为crea机房，三网回程CN2 GIA，带20G防御，常看我测评的朋友应该知道，一般带防御去程都是骨干线路，香港的线路也是CN2直连大陆，目前商家重新开业，价格非常美丽，性价比较非常高，有需要的朋友可以关注一下。活动方案...
CloudCone2核KVM美国洛杉矶MC机房机房2.89美元/月，美国洛杉矶MC机房KVM虚拟架构2核1.5G内存1Gbps带宽，国外便宜美国VPS七月特价优惠
近日CloudCone发布了七月的特价便宜优惠VPS云服务器产品，KVM虚拟架构，性价比最高的为2核心1.5G内存1Gbps带宽5TB月流量，2.89美元/月，稳定性还是非常不错的，有需要国外便宜VPS云服务器的朋友可以关注一下。CloudCone怎么样？CloudCone服务器好不好？CloudCone值不值得购买？CloudCone是一家成立于2017年的美国服务器提供商，国外实力大厂，自己开...
b2b程序为你推荐
-
香港iphone操作http重庆网络公司一九互联重庆本地的网约车平台有哪些？如何识别比较正规的网约车平台？德国iphone禁售令有人说苹果手机从2017年开始，中国禁售了asp.net什么叫ASP.NET?企业电子邮局企业邮箱怎么使用？全国企业信息查询有没有可以查全国企业信息的工商查询网站asp.net网页制作ASP.NET设计网页的方法？银花珠树晓来看晚来天欲雪,能饮一杯无。相似的句子我爱试用网我发现我对性爱这个话题好敏感！来吧看谁能把我下面说湿了？要200以上的才好评啊！