畢業(yè)設(shè)計外文文獻(xiàn)-基于改進(jìn)注意力循環(huán)控制門的品牌排序系統(tǒng)

附錄A外文翻譯—原文部分BrandrankingsystembasedonimprovedattentioncyclecontrolgateABrand-levelRankingSystemwiththeCustomizedAttention-GRUIne-commercewebsitessuchasTaobao,brandsareplayinganincreasinglyimportantroleinusers'click/purchaseselectionofproducts,inpartbecauseusersarenowmoreandmoreconcernedaboutthequalityofgoods,andbrandsareaqualityofgoods.Guarantee.However,theexistingsortingsystemisnotdesignedfortheuser'spreferenceforthebrand.Somesolutionscanalleviatethisproblemtoacertainextent,butstillcan'tachievethedesiredeffectorneedtoaddadditionalinteractioncosts.Weproposedanddesignedthefirstbrand-levelsortingsystemtosolvethisproblem.Thecorechallengeofthesystemishowtoeffectivelyusethelargeamountofinformationinthee-commercewebsitetopersonalizethebrand.Inoursolution,wefirstdesignfeatureengineeringforindividualizedbrandingissues,andthenbasedontheAttention-GRUmodel,threeimportantimprovementswereproposedtobetterrankthebrand.Itisworthnotingthattheseimprovementscanalsobeappliedtomanyothermachinelearningmodels.WedidalotofexperimentsonAlibabaTmallMalltoverifythevalidityofthissortingmodelandtestedthefeedbackfromusersonthesortingsysteminreal-worldscenarios.JUMP:AcollaborativepredictorforclickanddwelldurationInsearchandrecommendationscenarios,session-basedpredictionsareincreasinglybeingvaluedandvalued;mostofthesealgorithmsarebasedonrecurrentneuralnetwork(RNNs)techniquesorvariantsthereof.However,existingalgorithmsignoretheimportanceof"stayingtime"inexpressingthedegreeofuserpreference,andontheotherhand,intheveryshortornoisysession.Therefore,intheworkofthisarticle,weproposeajointestimationalgorithmJUMP,whichpredictstheuser'sclickandstaydurationbasedonthesession.JUMPusesanovelthree-layerRNNstructuretoencodeauser'ssession,includingusinga"fastandslowlayer"toalleviateshort-sessionproblemsanda"focuslayer"tosolvesessionnoiseproblems.AlargenumberofexperimentsshowthattheJUMPalgorithmcansurpassthelatestalgorithmsintermsofclickrateestimationandstaydurationestimation.E-commercesearchglobalrankingmethodThetraditionalmethodofsearchsortingistoscoretheproductsbyvariousmethods,andfinallysortbythescoreofeachitem.Thistraditionalsearchsortingmethodcannottakeintoaccountthemutualinfluenceofthedisplayedproducts.Similarly,thetraditionalmethodofestimatingctrandcvrforasinglecommodityisbasedontheassumptionthatthectrandcvrofthecommodityarenotaffectedbyothercommoditiesthataresimultaneouslydisplayed(wecallthepresentationcontext).Infact,thedisplaycontextofaproductcanaffecttheuser'sclickorpurchasedecision:Ifthegoodsaroundaproductaresimilartoit,butthepriceischeaperthanit,thentheprobabilityoftheuserbuyingitisnothigh;Thegoodsaremoreexpensivethanit,sotheprobabilityofusersbuyingitwillincrease.Ifyoubreaktheassumptionthatthetraditionalsortingmodelshowsthatthecontexthasnoeffect,howdoyousortit?Tothisend,wehaveforthefirsttimeproposedaglobalrankingmethodthatconsiderstheinteractionbetweencommodities.Wedescribethee-commercerankingasaglobaloptimizationproblem.Thegoalofoptimizationistoreflectthecustomerturnoverofthecustomersatisfaction:GMV(GrossMerchandiseVolume).Tobeprecise,thegoalofoptimizationforglobalorderingistomaximizethemathematicalexpectationsofGMV.CalculatingthemathematicalexpectationofGMVrequiresknowingtheprobabilityofthetransactionofthecommodity,andtheprobabilityofthetransactionofthecommodityinteractswitheachother.Therefore,weproposeatransactionprobabilityestimationmodelthatconsiderstheinteractionbetweencommodities.First,weproposeaconceptofglobalfeatureexpansion.Whenestimatingthetransactionprobabilityofacommodity,theinfluenceofothercommoditiesisaddedtotheprobabilityestimationmodelintheformofglobalfeatures,sothattheinfluenceofothercommoditiesistakenintoaccountintheestimation.Then,wefurtherproposetheRNNmodeltoaccuratelyconsidertheimpactoftheorderofthegoodsontheprobabilityofcommodityturnover.ByusingtheRNNmodel,weturnthee-commercesortingintoasequence-generatedproblemandusethebeamsearchalgorithmtofindabettersort.WeconductedalotofexperimentsontheTaobaowirelessmainsearchplatform.ComparedwiththeTaobaowirelessmainsearchalgorithmatthattime,weachieveda5%improvementinGMV.

附錄B外文翻譯—譯文部分基于改進(jìn)注意力循環(huán)控制門的品牌排序系統(tǒng)ABrand-levelRankingSystemwiththeCustomizedAttention-GRU在淘寶網(wǎng)等電子商務(wù)網(wǎng)站中,品牌在用戶對商品做點(diǎn)擊/購買選擇時正起著越來越重要的作用,部分原因是用戶現(xiàn)在越來越關(guān)注商品的質(zhì)量,而品牌是商品質(zhì)量的一個保證。但是,現(xiàn)有的排序系統(tǒng)并不是針對用戶對品牌的偏好設(shè)計的。某些處理方案一定程度上能減輕這個問題,但仍然無法取得理想的效果或需要增加額外的交互成本。我們提出并設(shè)計了第一個品牌級排序系統(tǒng)來解決該問題,該系統(tǒng)的核心挑戰(zhàn)是如何有效利用電子商務(wù)網(wǎng)站中的大量信息來對品牌作個性化排序。在我們的解決方案中,我們首先針對個性化品牌排序問題設(shè)計特征工程,然后在Attention-GRU模型的基礎(chǔ)上,提出了3個重要改進(jìn),以更好地對品牌排序。值得注意的是,這些改進(jìn)也能應(yīng)用于很多其他機(jī)器學(xué)習(xí)模型。我們在阿里巴巴天貓商城上做了大量實(shí)驗(yàn)來驗(yàn)證這個排序模型的有效性,并測試了真實(shí)應(yīng)用場景中用戶對該排序系統(tǒng)的反饋。JUMP:一種點(diǎn)擊和停留時長的協(xié)同預(yù)估器在搜索和推薦場景中，基于會話的預(yù)測日益受到人們的興趣和重視；大多這類算法都是基于遞歸神經(jīng)網(wǎng)絡(luò)(RNNs)技術(shù)或者它的變種。然而，已有的算法一方面會忽略「停留時長」在表達(dá)用戶偏好程度上的重要性，另一方面在非常短或者有噪音的會話上會預(yù)估的不準(zhǔn)。因此在該文章的的工作中，我們提出了一種聯(lián)合預(yù)估算法JUMP，基于會話去同時預(yù)測用戶的點(diǎn)擊和停留時長。JUMP使用一種新奇的三層RNN結(jié)構(gòu)去編碼用戶的一次會話，包括使用一個「快慢層」去緩解短會話的問題，使用一個「注意力層」去解決會話噪音的問題。大量的實(shí)驗(yàn)表明JUMP算法能在點(diǎn)擊率預(yù)估和停留時長預(yù)估上，超越最新的其他算法。電商搜索全局排序方法搜索排序的傳統(tǒng)方法是通過各種方法對商品進(jìn)行打分，最后按照每個商品的分?jǐn)?shù)進(jìn)行排序。這樣傳統(tǒng)的搜索排序方法就無法考慮到展示出來的商品之間相互的影響。類似地，傳統(tǒng)的針對單個商品估計ctr、cvr的方