




版权说明:本文档由用户提供并上传,收益归属内容提供方,若内容存在侵权,请进行举报或认领
文档简介
1、CS 345Data MiningOnline algorithmsSearch advertisingOnlinealgorithmsClassic model of algorithmsYouget to seetheentire input,then computesomefunctionofitInthis context, “offline algorithm”OnlinealgorithmYouget to seetheinputone piece at atime,and needtomakeirrevocabledecisionsalongthe waySimilar to d
2、atastream modelsExample:Bipartitematching1234abcdGirlsBoysExample:Bipartitematching1234abcdM =(1,a),(2,b),(3,d)isamatchingCardinalityofmatching= |M|= 3GirlsBoysExample:Bipartitematching1234abcdGirlsBoysM =(1,c),(2,b),(3,d),(4,a)isaperfect matchingMatchingAlgorithmProblem:Find amaximum-cardinalitymat
3、chingforagivenbipartitegraphA perfectone if it existsThereisa polynomial-timeoffline algorithm(Hopcroftand Karp1973)Butwhatifwedonthave theentiregraphupfront?OnlineproblemInitially, we aregiventheset BoysIneach round,onegirls choicesare revealedAtthat time, we havetodecide to either:Pair thegirl wit
4、haboyDontpair thegirl withany boyExample of application:assigningtaskstoserversOnlineproblem1234abcd(1,a)(2,b)(3,d)GreedyalgorithmPair thenewgirlwith anyeligibleboyIfthereisnone,dontpair girlHowgoodisthealgorithm?CompetitiveRatioForinputI,supposegreedyproducesmatchingMgreedywhileanoptimal matching i
5、s MoptCompetitiveratio=minallpossibleinputs I(|Mgreedy|/|Mopt|)Analyzingthe greedyalgorithmConsidertheset GofgirlsmatchedinMoptbutnot in MgreedyThen it mustbethe casethateveryboyadjacenttogirlsinGisalready matchedinMgreedyTheremust be at least |G|such boysOtherwisethe optimalalgorithm could nothave
6、matchedall theG girlsTherefore|Mgreedy|G|=|Mopt- Mgreedy|Mgreedy|/|Mopt|1/2Worst-case scenario1234abc(1,a)(2,b)dHistory of webadvertisingBannerads (1995-2001)Initial formofweb advertisingPopular websites chargedX$for every 1000“impressions”ofadCalled“CPM”rateModeled similartoTV, magazine adsUntarget
7、ed to demographicallytagetedLowclickthrough rateslowROI foradvertisersPerformance-basedadvertisingIntroduced by Overture around2000Advertisers“bid”onsearchkeywordsWhen someonesearchesfor thatkeyword,the highestbiddersadisshownAdvertiser is chargedonlyiftheadisclickedonSimilar model adoptedbyGoogle w
8、ithsomechanges around2002Called“Adwords”Adsvs. searchresultsWeb2.0Performance-basedadvertisingworks!Multi-billion-dollarindustryInterestingproblemsWhat adstoshow fora search?IfImanadvertiser,whichsearch terms shouldI bidonandhow muchtobid?Adwords problemA streamofqueries arrivesatthe searchengineq1,
9、q2,Several advertisers bidoneach queryWhen query qiarrives,searchengine mustpicka subsetofadvertiserswhoseadsare shownGoal:maximizesearchenginesrevenuesClearly we needanonline algorithm!GreedyalgorithmSimplestalgorithmisgreedyIts easytosee thatthe greedyalgorithmisactuallyoptimal!Complications (1)Ea
10、ch ad hasa differentlikelihood of being clickedAdvertiser 1bids$2,clickprobability=0.1Advertiser 2bids$1,clickprobability=0.5ClickthroughratemeasuredhistoricallySimplesolutionInstead of rawbids,usethe “expectedrevenue perclick”TheAdwordsInnovationAdvertiserBidCTRBid*CTRABC$1.00$0.75$0.501%2%2.5%1 ce
11、nt1.5cents1.125centsTheAdwordsInnovationAdvertiserBidCTRBid*CTRABC$1.00$0.75$0.501%2%2.5%1 cent1.5cents1.125centsComplications (2)Each advertiserhas alimitedbudgetSearchengine guaranteesthattheadvertiserwill notbecharged morethantheirdailybudgetSimplified model (fornow)Assumeall bidsare 0or1Each adv
12、ertiserhas thesame budgetBOneadvertiserperqueryLetstrythe greedyalgorithmArbitrarilypick an eligible advertiserfor eachkeywordBadscenariofor greedyTwoadvertisersAandBA bidsonqueryx,Bbids on xand yBoth havebudgetsof$4Querystream: xxxxyyyyWorstcase greedychoice: BBBB_Optimal:AAAABBBBCompetitiveratio=
13、SimpleanalysisshowsthisistheworstcaseBALANCE algorithmMSVVMehta, Saberi,Vazirani, andVaziraniForeachquery,picktheadvertiserwith thelargest unspentbudgetBreakties arbitrarilyExample:BALANCETwoadvertisersAandBA bidsonqueryx,Bbids on xand yBoth havebudgetsof$4Querystream: xxxxyyyyBALANCE choice:ABABBB_
14、Optimal:AAAABBBBCompetitiveratio= AnalyzingBALANCEConsidersimplecase:two advertisers,A1andA2, eachwithbudgetB(assume B1)Assumeoptimalsolutionexhaustsboth advertisersbudgetsBALANCE mustexhaustatleastoneadvertisers budgetIfnot, we canallocatemore queriesAssumeBALANCEexhaustsA2sbudgetAnalyzingBalanceA1
15、A2BxyBA1A2xOptrevenue= 2BBalance revenue=2B-x =B+yWehave yxBalance revenueisminimumforx=y=B/2Minimum Balancerevenue= 3B/2CompetitiveRatio= 3/4Queries allocatedtoA1inoptimal solutionQueries allocatedtoA2inoptimal solutionGeneral ResultInthegeneralcase,worstcompetitiveratioofBALANCE is11/e= approx.0.6
16、3Interestingly,noonlinealgorithm hasa bettercompetitiveratioWontgothrough thedetails here, butletsseethe worst casethatgivesthis ratioWorstcase forBALANCEN advertisers,each withbudget BN1NBqueries appearinN roundsofB querieseachRound1 queries: biddersA1, A2, , ANRound2 queries: biddersA2, A3, , ANRo
17、undi queries: biddersAi, , ANOptimum allocation: allocate round iqueriestoAiOptimum revenueNBBALANCE allocationA1A2A3AN-1ANB/NB/(N-1)B/(N-2)Afterk rounds,sum of allocations to eachofbinsAk,ANisSk= Sk+1= =SN=11kB/(N-i+1)Ifwefind thesmallestk suchthatSkB,then after kroundswecannotallocateany queriesto
18、any advertiserBALANCE analysisB/1B/2B/3B/(N-k+1) B/(N-1)B/NS1S2Sk= B1/1 1/2 1/3 1/(N-k+1) 1/(N-1) 1/NS1S2Sk = 1 BALANCE analysisFact:Hn=1in1/i=approx. log(n)forlargenResultdue to Euler1/11/21/31/(N-k+1) 1/(N-1)1/NSk = 1 log(N)log(N)-1Sk= 1impliesHN-k= log(N)-1 =log(N/e)N-k=N/ek =N(1-1/e)BALANCE analysisSoafterthefirstN(1-1/e)rounds,wecannotallocateaquerytoanyadvertiserRevenue =BN(1-1/e)Competitiveratio= 1-1/eGeneral versionofproblemArbitrarybids,budgetsConsiderqueryq,advertiser iBid=xiBudget=biBALANCE canbeter
温馨提示
- 1. 本站所有资源如无特殊说明,都需要本地电脑安装OFFICE2007和PDF阅读器。图纸软件为CAD,CAXA,PROE,UG,SolidWorks等.压缩文件请下载最新的WinRAR软件解压。
- 2. 本站的文档不包含任何第三方提供的附件图纸等,如果需要附件,请联系上传者。文件的所有权益归上传用户所有。
- 3. 本站RAR压缩包中若带图纸,网页内容里面会有图纸预览,若没有图纸预览就没有图纸。
- 4. 未经权益所有人同意不得将文件中的内容挪作商业或盈利用途。
- 5. 人人文库网仅提供信息存储空间,仅对用户上传内容的表现方式做保护处理,对用户上传分享的文档内容本身不做任何修改或编辑,并不能对任何下载内容负责。
- 6. 下载文件中如有侵权或不适当内容,请与我们联系,我们立即纠正。
- 7. 本站不保证下载资源的准确性、安全性和完整性, 同时也不承担用户因使用这些下载资源对自己和他人造成任何形式的伤害或损失。
最新文档
- 子宫肌瘤诊治规范
- 教师基层服务工作总结
- 阿坝师范学院《大学基础听说4》2023-2024学年第二学期期末试卷
- 扫黑除恶专项斗争工作培训
- 陕西交通职业技术学院《数学竞赛指导》2023-2024学年第一学期期末试卷
- 陕西机电职业技术学院《国际会展实务》2023-2024学年第二学期期末试卷
- 陕西电子信息职业技术学院《舆情监测与危机应对》2023-2024学年第二学期期末试卷
- 国家电子商务标准体系研究
- 陕西省榆林市绥德县2025届初三化学试题模拟卷试题含解析
- 陕西省汉中市部分高中2025年高三下学期5月月考生物试题文试题含解析
- 雷锋叔叔你在哪里教学反思
- (新版)国家统计执法证资格考试备考题库(含答案)
- 项目验收单标准模板
- 24式太极拳教案(1~4课)
- 小学 三年级 心理健康《最好的老师-兴趣的作用》教学设计
- DB12T 1040-2021 建筑工程规划管理技术规范
- 中国绿色经济发展之路(PPT-37张)课件
- G322-1钢筋砼过梁
- 客房控制系统——RCU系统培训PPT通用通用课件
- 压力管道安装许可证换证自评报告
- (会议纪要(2011)第29期)河南煤业化工集团有限责任公司会议纪要
评论
0/150
提交评论