人工智能KDD2022InterpretationofselectedpapersbyMeituantechnicalteam美团技术团队

637 473

|Theauthorsofthepaper:LiuXiao(TsinghuaUniversity),ZhaoShiyu(TsinghuaUniversity),SuKai(TsinghuaUniversity),CenYukuo(Meituan),QiuJiezhong(TsinghuaUniversity),DongYuxiao(TsinghuaUniversity),ZhangMengdi(Meituan),Wuwei(Meituan),TangJie(TsinghuaUniversity)

|PaperIntroduction:KnowledgeGraphPre-trainingforComplexLogicalQuery.Thispaperstudiesthecomplexlogicalqueryprobleminknowledgegraph,discussestheinherentdefectsofmainstreamreasonersbasedonknowledgegraphembedding,andproposesanewgraphneuralnetworkreasonerbasedonKGTransformer,anditscorrespondingpre-trainingandfine-tuningmethods.KGTransformerachievesthebestresultsontwomajorknowledgegraphreasoningdatasets,especiallyonout-of-domaintasks,whichprovesthebroadprospectsofthisideaforknowledgegraphreasoning.

|Introductiontothepaper:Theindustrial-levelsearchrecommendationsystemmainlyfollowsthealgorithmsystemofrecall,roughsorting,finesorting,andrearranging.Inordertomeetthehugescoringscaleandstrictdelayrequirementsofroughrow,thetwin-towermodelisstillwidelyused.Inordertoimprovetheeffectofthemodel,someschemeswilladditionallyusetherefinedscoringknowledgefordistillation.However,twomajorchallengesremaintobeaddressed:

ThispaperusesthemethodofNeuralArchitectureSearch,andinnovativelyproposesthealgorithmframeworkofAutoFAS(AutomaticFeatureandArchitectureSelectionforPre-RankingSystem).Undertheguidanceoftheknowledgeoffinesortingandscoring,theoptimalcombinationschemeofroughsortingfeaturesandstructureisselectedatthesametime,whichachievestheeffectofSOTA.ThissolutionhasbeenfullyusedinthemainsearchscenarioofMeituan,andhasachievedsignificantonlineimprovement.

|Theauthorsofthepaper:LiYinfeng(TsinghuaUniversity),GaoChen(TsinghuaUniversity),DuXiaoyi(Meituan),WeiHuazhou(Meituan),LuoHengliang(Meituan),JinDepeng(TsinghuaUniversity),LiYong(TsinghuaUniversity)

Specifically,inthedisentangledintentencoder,weconstructthreesetsofdualhypergraphstocapturehigher-orderrelationshipsunderthreedifferenttypesofpreferences(time-related,place-related,andintrinsic)Theinformationdisseminationmechanismonthenetworklearnsdisentangledintentrepresentationsforusers.Intheintentdiscoverydecoder,thispaperconstructsthepseudo-labelsofpairedsamplesbasedonthesimilarityofdenoisedintentrepresentations,andrealizestheknowledgetransferfromknownintentstounknownintentsthroughsemi-supervisedlearningtocompleteintentdiscovery.Thispapercompareswithavarietyofadvancedbenchmarkalgorithmsonthelarge-scaleindustrialdatasetofMeituan.TheexperimentalresultsshowthattheproposedAutoIntentmethodcanachieveasignificantperformanceimprovementofmorethan15%comparedwiththeexistingbestsolutions.Overall,thispaperprovidesanewresearchideaforunderstandingandmodelinguserconsumptionbehaviorincities.

|Theauthorsofthepaper:LiuChang(TsinghuaUniversity),YuanYuan(TsinghuaUniversity),GaoChen(TsinghuaUniversity),BaiChen(Meituan),LuoLingrui(Meituan),DuXiaoyi(Meituan),ShiXinlei(Meituan),LuoHengliang(Meituan),JinDepeng(TsinghuaUniversity),LiYong(TsinghuaUniversity)

|Theauthorsofthepaper:LiuYaxu(NationalTaiwanUniversity&Meituanintern),YanRuinan(NationalTaiwanUniversity),YuanBowen(NationalTaiwanUniversity&Meituanintern),ShiRundong(Meituan),YanPeng(Meituan),LinZhiren(TaiwantheUniversity)

|PaperIntroduction:Fortrainingamachinelearningmodel,akeytaskistoconstructtrainingdatafromthefeedbackcollected(e.g.,ratings,clicks).However,itcanbefoundfromtheoreticalandpracticalexperiencethattheselectionbiasinthecollectedfeedbackcanleadtoabiasedmodelobtainedbytraining,resultinginthetrainingresultbeingnottheoptimalpolicy.Tosolvethisproblem,counterfactuallearninghasreceivedalotofattention,andexistingcounterfactuallearningmethodscanbedividedintoValueLearningmethodsandPolicyLearningmethods.

ThispaperstudiesthePolicyLearningmethodoftheTop-rankingmodelwithalargerdecisionspace,andproposesapracticallearningframeworktosolvetheImportanceWeightexplosioninthelearningofalargerdecisionspace,lesssamplesleadtolargervariance,andtraininglowefficiency,etc.Experimentsonopensourcedataverifytheeffectivenessandefficiencyoftheproposedframework.

|Theauthorsofthepaper:GaoChengliang(Meituan),ZhangFan(Meituan),ZhouYue(Meituan),FengRonggen(Meituan),RuQiang(Meituan),BianKaigui(PekingUniversity),HeRenqing(Meituan)Mission),SunZhizhao(Meituan)

|Introductiontothepaper:Intheinstantdeliverysystem,theaccurateestimationoftheordertimeforthemerchanttodeliverthemealisveryvaluabletoboththeuserandtheriderexperience.Therearetwomaintechnicalchallengesinthisproblem,namely,incompletesamplelabels(someordersonlyhavetheapproximaterangeofmealtime)andlargedatauncertainty,whicharedifficulttodealwithbyconventionalpointestimationregressionmethods.

Inthiswork,probabilityestimationisusedforthefirsttimetocharacterizetheuncertaintyofordermealtime,andanon-parametricmethodbasedondeeplearningisproposed,andthedatasamplesofrangelabelsarefullyutilizedinfeatureconstructionandmodeldesign.Inprobabilityestimation,thispaperproposestheS-QLlossfunctionandprovesitsmathematicalrelationshipwithS-CRPS,basedonwhichthequantilediscretizationofS-CRPSisperformedtooptimizethemodelparameters.BasedonrealdeliverydataevaluationandonlineA/Bexperiments,theadvantagesandeffectivenessofthismethodhavebeenproved.

|Theauthorsofthepaper:WuZhuolin(Meituan),HuangFangsheng(Meituan),ZhouLinjun(Meituan),SongYu(Meituan),YeChengpeng(Meituan),NiePengyu(Meituan),RenHao(Meituan),HaoJinghua(Meituan)Meituan),HeRenqing(Meituan),SunZhizhao(Meituan)

|Introductiontothepaper:Meituandeliveryaimstoprovidecustomersandrestaurantswithhigh-qualityandstableservices,buthundredsofthousandsofordersarestillcancelledeverydaybecausenoonetakestheorder.ThecancellationofordershascausedgreatdamagetotheuserexperienceandthereputationofMeituandamage.Tosolvethisproblem,Meituanprovidedaspecialfundtoimprovetheuserexperienceoftailorders.Ordersthathavenotbeenpickedupwillcontinuetobeexposedtoriders,soweneedtocontinuouslydecideontheamountofadditionalrewardsforordersuntiltheorderiscancelledortaken.Sincetheincentiveschemeatthelastmomentoftheorderwillsignificantlyaffecttheprobabilityoftheexistenceandcancellationoftheorderinthesubsequentstages,thisproblemisacomplexmulti-stagesequentialdecision-makingproblem.Inordertobetterenhancetheuserexperience,weproposeanewframeworktosolvethisproblem.Thisframeworkconsistsofthreeparts:

Thesemi-parameterizedordercompletion(cancellation)probabilitymodelisusedtopredicttherelationshipbetweentherewardamountallocatedtotheorderandtheprobabilityoftheorderbeingpickedupandfinallycompleted(cancelled)atthismoment.TheLagrangiandualdynamicprogrammingalgorithmmainlyusesThehistoricalorderdatacalculatestheLagrangianmultipliersolutionforeachallocationsequence,andtheonlineallocationalgorithmusestheresultsobtainedintheofflineparttocalculatethecorrespondingincentiveschemeforeachorder.WeconductedA/Bexperimentsonrealdeliveryscenarios.Theexperimentalresultsshowthatthenewalgorithmreducesthenumberofcancelledordersby25%comparedwiththebaselinealgorithm,whichsignificantlyimprovestheuserexperience.

TheabovepapersaretheresultsofthejointcooperationbetweenthetechnicalteamofMeituanandvariousuniversitiesandscientificresearchinstitutions.ThispapermainlyintroducessomescientificresearchworkofMeituaninthetechnicalfieldsofgraphpre-training,selectionalgorithm,automaticintentiondiscovery,effectmodeling,policylearning,probabilityprediction,andrewardframework.Ihopeitcanbehelpfulorenlighteningtoyou,andyouarealsowelcometocommunicatewithus.

ReadmorecollectionsoftechnicalarticlesfromtheMeituantechnicalteam

|Replykeywordssuchas[2021stock],[2020stock],[2019stock],[2018stock],[2017stock]inthepublicaccountmenubardialogbox,youcanviewthecollectionoftechnicalarticlesbytheMeituantechnicalteamovertheyears.

THE END

KD最新信息，引领潮流，洞悉未来发展趋势，KD最新动态，引领潮流，洞悉未来发展趋势推荐

KD最新信息详解，从数据到趋势，全方位解读KD的未来走向，KD最新情报解析，数据趋势与未来展望澳门

KDD2022重磅奖项出炉，刘欢获创新奖，耶鲁清华和慕尼黑理工博士摘得博士论文奖

KDD2022探究图神经网络GNN过平滑的真正原因

人工智能KDD2022InterpretationofselectedpapersbyMeituantechnicalteam美团技术团队

1.比乙欧本vsKMSK丹泽伤病名单最新(2024年12月15日)2022-11-09 KMSK丹泽(3 : 0)欧本 2018-12-06 KMSK丹泽(1 : 3)欧本 2015-12-13 KMSK丹泽(1 : 1)欧本 2015-08-13 欧本(5 : 1)KMSK丹泽 2009-02-22 KMSK丹泽(1 : 2)欧本 2008-10-05 欧本(2 : 2)KMSK丹泽 欧本近6主场取得1胜0平5负,进7球失16球,胜率17%;KMSK丹泽近6客场取得0胜2https://www.qiumiwu.com/news/1522848480914
2.KDDConferenceKDD: Knowledge Discovery and Data Mining KDD '23: Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining KDD: Knowledge Discovery and Data Mining ASONAM '22: Proceedings of the 2022 IEEE/ACM International Conference on Advances in Social Networks Analysis and Mining https://dl.acm.org/conference/kdd
3.KDD2022WashingtonDC,U.S.He also received the 2009 Fulkerson Prize, 2021 ACM STOC Test of Time Award (for smoothed analysis), 2022 ACM SIGecom Test of Time Award (for settling the complexity of computing a Nash equilibrium), and 2011 ACM STOC Best Paper Award (for improving maximum-flow minimum-cut algorithms). https://www.kdd.org/kdd2022/keynoteSpeaker.html
4.dblp:HPCC/DSS/SmartCity/DependSys202224th IEEE Int Conf on High Performance Computing & Communications; 8th Int Conf on Data Science & Systems; 20th Int Conf on Smart City; 8th Int Conf on Dependability in Sensor, Cloud & Big Data Systems & Application, HPCC/DSS/SmartCity/DependSys 2022, Hainan, China, December 18-20, 2022https://dblp.org/db/conf/hpcc/hpcc2022
5.kddilogo;kddilogo?(2022年10月11日更新)当社や公的機関などを装う偽SMSにご注意ください ?Web請求書の過去請求書参照期間延長について ?KDDI利用料金の「未払い金」やau PAYなど当社グループサービスを装う偽メールにご注意ください ?宅配業者や金融機関、携帯電話会社など、企業からのお知らせを装って発https://www.kddi.com/
6.KDD2022All KDD attendees were invited to stop by the Apple booth (booth number 144) to chat with available recruiters and booth staff. Schedule Below was the schedule of Apple events at KDD 2022. Sunday August 14 WORKSHOP 2nd International Workshop on Online and Adaptive Recommender Systems https://machinelearning.apple.com/updates/apple-at-kdd-2022
7.PaperDigest:KDD2022Highlights–PaperDigestPaper Digest Team New York City, New York, 10017 team@paperdigest.org TABLE 1: Paper Digest: KDD 2022 Highlights : In this paper, we propose FedSVD, a practical lossless federated SVD method over billion-scale data, which can simultaneously achieve lossless accuracy and high efficiency. https://www.paperdigest.org/2022/08/kdd-2022-highlights/
8.KDD2022美团技术团队精选论文解读今年,美团技术团队有多篇论文被KDD 2022收录,这些论文涵盖了图谱预训练、选择算法、意图自动发现、效果建模、策略学习、概率预测、奖励框架等多个技术领域。本文精选了7篇论文做简要介绍(附下载链接,论文排名不分先后),希望能对从事相关研究方向的同学有所帮助或启发。 https://blog.51cto.com/u_15197658/5686667
9.干货!KDD2022博士论文奖冠亚军对话KDD2022 博士论文奖获奖结果已经公布,来自斯坦福大学的Rex Ying(应智韬)获得博士论文奖WINNER 奖,清华大学裘捷中获得Runner Up 奖。 KDD,全称国际数据挖掘与知识发现大会,是数据挖掘领域历史最悠久、规模最大的国际顶级学术会议,也是首个引入大数据、数据科学、预测分析等概念的会议。会议自1995年举办第一届以来,今年的https://blog.csdn.net/AITIME_HY/article/details/126535354
10.KDD2022August 14, 2022–August 18, 2022 Microsoft at KDD 2022 Eastern Daylight Time (UTC -4) Location:Washington, DC Microsoft is proud to be a sponsor of the28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining(opens in new tab). The annual ACM SIGKDD conference is the premier intehttps://www.microsoft.com/en-us/research/event/kdd-2022/
11.京东科技郑宇获KDD2022时间检验应用数据科学奖鞭牛士报道 近日,KDD 2022的时间检验奖、博士论文奖等重磅奖项陆续揭晓,袁晶、郑宇、谢幸等人2012年的论文Discovering Regions of Different Functions in a City Using Human Mobility and POls获得大会时间检验奖-应用数据科学方向(SIGKDD Test of Time Award Applied Science)。 http://www.bianews.com/news/details?id=135478
12.计算机系教师东昱晓获得SIGKDD2022新星奖计算机系教师东昱晓获得SIGKDD2022新星奖 8月16日,国际数据挖掘与知识发现大会(SIGKDD)官方公布了2022年ACM SIGKDD(简称 KDD)创新奖、服务奖、新星奖、博士论文奖、时间检验研究奖、时间检验应用科学奖六大奖项的获得者。计算机系教师东昱晓获得SIGKDD2022新星奖(SIGKDD Rising Star Award),以表彰他在博士毕业5年中https://www.cs.tsinghua.edu.cn/info/1034/5577.htm
13.阿里巴巴达摩院团队独立获得KDD2022大奖“最佳论文奖”KDD组委会评价,这一工作“推动了联邦图学习技术的发展,并树立了优秀平台工作的榜样”。今年以来,阿里巴巴在全球顶级学术会议上多次荣获大奖,覆盖自然语言处理、计算机视觉、数据科学等领域最高会议,囊括KDD 2022最佳论文、ACL 2022杰出论文、CVPR 2022最佳学生论文等奖项。https://www.elecfans.com/d/1881801.html
14.计算机学院商烁教授团队研究生在数据挖掘顶会KDD2022发表研究5月19日,数据挖掘领域顶级会议KDD 2022 28th SIGKDD Conference on Knowledge Discovery and Data Mining)公布论文录取结果,我校计算机学院时空大数据与智能团队2020级硕士研究生饶漩、2021级硕士研究生刘大川分别以第一作者撰写的论文被接收,届时他们将在会议上展示成果并做学术报告。 https://www.1633.com/article_univ/233389.html
15.ConferenceonKnowledgeDiscoveryandDataMiningACM SIGKDD International Conference on Knowledge Discovery and Data Mining - KDD 2024 ACM SIGKDD International Conference on Knowledge Discovery and Data Mining - KDD 2023 ACM SIGKDD International Conference on Knowledge Discovery and Data Mining - KDD 2022 https://www.ourglocal.com/wikicfp/?conid=37&year=2022
16.KDD2022中科院计算所提出STABLE:一种无监督高鲁棒性图结构近年来,得益于消息传递机制,图神经网络(GNN)在大量基于图数据的任务上取得了卓越的成效,尤其是半监督节点分类任务。然而,最新的研究表明,GNN在对抗攻击面前是极为脆弱的,尤其是结构扰动。攻击者只需要对图结构进行微小的篡改就能使的模型https://mp.weixin.qq.com/s?__biz=MzAxMzc2NDAxOQ==&mid=2650432298&idx=2&sn=27559e930d0112abf5c98be5c102c163&chksm=839357f6b4e4dee00d6ddc9bd3d9977d90c2709fb656cfaa68df47116bc9bec5029d7d656fd9&scene=27
17.Q/KDDY0003S2022巴氏杀菌蛋液系列食品本标准规定了巴氏杀菌蛋液系列的分类、要求、检验规则和标志、包装、运输、贮存及保质期。本标准适用于以鲜蛋为原料,经预处理、打蛋、分离或不分离、搅拌过滤、均质、巴氏杀菌、包装而成的食品加工用巴氏杀菌蛋液系列(以下简称产品)。http://down.foodmate.net/wap/index.php?itemid=136689&moduleid=23
18.增量/持续学习2022KDDStreamingGraphNeuralNetworksvia【增量/持续学习】2022 KDD Streaming Graph Neural Networks via Generative Replay 目录 1、简介 2、Intro 3、SGNN-GR 动机 贡献 相关工作 基础 方法https://www.1000bd.com/Article/Index/821860
19.GitHubmain BranchesTags Code Folders and files Name Last commit message Last commit date Latest commit History 27 Commits KDD22 examples fdsa .gitignore LICENSE README.md conda.yml requirements.txt setup.py README MIT license A fully differentiable set autoencoder for encoding sets.Paper @KDD 2022. https://github.com/PaccMann/fdsa
20.尤克达蒂(yookdd)奥特曼衣服男童2022新款cosplay超人小孩超人梦比优尤克达蒂(yookdd)奥特曼衣服男童2022新款cosplay超人小孩超人梦比优斯连体衣 梦比优斯单连体衣 110cm图片、价格、品牌样样齐全!【京东正品行货,全国配送,心动不如行动,立即购买享受更多优惠哦!】https://item.jd.com/10128037850942.html
21.2022年张家口开放大学校长办公室及党委办公室 2022年11月10日 1,178 views 0 迎接党的二十大 培根铸魂育新人—张家口开放大学召开庆祝第38个教师节暨表彰大会 秋澄景清,硕果馥郁。 9月9日下午,张家口开放大学隆重举行以“迎接党的二十大,培根铸魂育新人”为主题的庆祝第三十八个教师节表彰大会。学校党委书记杨汝明、校长张宏文,校领 .http://zjkdd.com.cn/2022/
22.北京大学数据与智能实验室(PKU2022年05月回顾 【论文录用】 2022年05月,实验室共有8篇论文被国际顶级学术会议或期刊录用: ① 我组博士生张文涛、盛则昂、沈彧等同学合作的两篇论文被ICML 2022录用; ② 我组博士生符芳诚、苗旭鹏的一篇论文被VLDB 2022录用; ③ 我组博士生黎洋、沈彧等同学合作的两篇论文被KDD 2022录用; https://www.scholat.com/org/orgPost.html?id=51