Propaga'on)based.Social)Aware.Replica'on. for.Social.Video.Contents. - PowerPoint PPT Presentation

Propaga'on)based.Social)Aware.Replica'on. for.Social.Video.Contents. Zhi$Wang,$Lifeng$Sun,$Xiangwen$Chen,$Wenwu$Zhu,$$ Jiangchuan$Liu,$Minghua$Chen,$and$Shiqiang$Yang$ Tsinghua)University,)China) The)Chinese)University)of)Hong)Kong) Simon)Fraser)University,)Canada) ACM Multimedia 2012, Oct. 29 – Nov. 2, Nara, Japan

12K $re9shares$ 5K $re9shares$ 2K $re9shares$

Social.Network.Changes.How.People.Consume.Videos . Over$ 700 $YouTube$video$ links$are$imported$to$TwiEer$ Video& every$minute$ sharing & Over$ 500 $years$of$YouTube$ videos$are$watched$every$ day$on$Facebook$ Over$ 30%. users$select$videos$ using$social$network$service$ 80%. worldwide . growth$ (Q4)2011)@)Q1)2012)) in$China$ Social$network$is$fundamentally$changing$ video$distribuHon$landscape$

Social.Video.Distribu'on.Landscape.Change . TradiBonal)Video)DistribuBon � Social)Video)DistribuBon � Dynamically$ influenced$by$users$ PropagaHon$in$small$ social$groups$ • Based$on$popularity • Based$on$propagaHon$ Rethinking$distribuHon$for$social$video$contents$

Related.Works . Propaga'on) based.social) Content.correla'on.for. aware.replica'on.............. UGC.distribu'on. [INFOCOM’09] � [MM’12] � User.rela'on/influence.for. social.media.distribu'on. Cache,.CDN,.and.P2P. [SIGCOMM’10,$INFOCOM’12] � [TMM’04,$TMM’07,$MM’09] �

PSAR:.Propaga'on)based.Social)Aware.Replica'on. Propaga'on.Inference :$Jointly$consider$user,$content,$and$context$ $ Propaga'on.inference. $ PropagaBon)predicBon) Influence$ Region$predictor$ User) Global9audience$ Replica'on.based.on. $ predictor$ social.propaga'on. RelaHon � Local9audience$ predictor$ Preference$ Locality$ Strategy) Content) Context) $ $ PropagaBon)paKern) Architecture) Popularity � LocaHon � Social$locality$ Geo$locality$ Temporal$locality$

Data)driven.Approach:.Measurement.Study . Tencent$Weibo$ ! 0.5B. users$–$one$of$the$largest$social$network$service$in$China$ $ $ $ A$sample$of$video$sharing$in$ 20 $days$ ! 350K. videos$–$ 5 $video$sharing$websites,$ 14 $categories$ ! 1M $users$–$social$relaHon,$behaviors$ ! Over$ 100 $city9level$regions$–$locaHon,$distance$

Social.Locality . How$far$a$video ) propagates$along$social$relaHon ) • PropagaHon$links$are$ short$(average$length$ 1.21 )$ • Users$are$socially$ close$to$each$other$

Geographical.Locality . How$many$regions$a$video$propagates$to$ Most$videos$propagate$ • in$a$few$regions$$$$$$ ( 90% $<$ 5 $regions)$ Users$are$located$ • nearby $

Temporal.Locality . How$long$a$video$propagates$ • Re9shares$happen$in$ the$early$hours$$$$$$$$$ ( 95% $<$ 1 $day)$ • User$behaviors$are$ grouped$around$the$ import$Hme$

Architecture.Design.of.PSAR . PropagaBon)paKern :$social,$geographical$and$temporal$localiHes$ Content$propagaHon$ predicaHon $ � PropagaHon9based$ social9aware$replicaHon$ Edge9cloud$and$peer9 assisted$architecture$ �

Region.Predic'on . Predict$new$regions$a$video$propagates$to$ g ( T +1) = c 1 log( c 2 s ( T ) " Number$of$regions$involved:$ ) v v PropagaBon)size) " New$region:$ Friends)of)current)users) Number of regions in the propagation (y) 50 40 30 20 10 (x, y) (x, 6.5log(0.5682 x)) 0 0 1000 2000 3000 Propagation size (x)

Global)Audience.Predic'on . Predict$global$audience$aEracted$ Popularity) Temporal)locality) Social)locality) " Global$audience:$ e ( T +1) = s ( T ) ( z s ( τ ( T ) ) /h ( T ) ) v v v v $ corrcoef=0.6681 corrcoef=0.1628 3 3 10 10 New users (resource) New users (resource) 2 2 10 10 1 1 10 10 0 0 10 10 0 1 2 3 4 − 4 − 2 0 2 10 10 10 10 10 10 10 10 10 s v z(t)/h v (popularity+social) s v (popularity)

Local)Audience.Predic'on . Predict$local$friends$aEracted$ Local)popularity) " Local$audience:$ Friend)preference) 1 0.8 0.6 CDF 0.4 0.2 0 − 0.2 0 0.2 0.4 0.6 0.8 1 Preference correlation (T+1 vs. T)

Experiment.Setup . Trace$ ! A$sample$of$video$sharing$in$10$days$ ! Guideline:$videos$are$selected$with$diverse$replicaHon$$$$$$ and$replacement$ $ EvaluaHon$metric $ ! Local$download$raHo$ ! Local$cache$hit$raHo$ $

72%.Users.Locate.Their.Videos.in.2.Hops.in.PSAR . EffecHveness$of$the$architecture$ 80 PSAR Fraction of P2P requests (%) LFU − based Replacement 2$Hmes$improvement$as$compared$to$ Random Replacement 60 convenHonal$wisdoms$ 40 20 0 1 hop 2 hops others Number of social hops

Popular.and.Unpopular.Videos.Effec'vely.Replicated . EffecHveness$of$the$strategy$ 1 Local download ratio 0.8 0.6 4$Hmes$improvement$in$local$fetch$of$ 0.4 unpopular$videos$$ PSAR 0.2 Popuarlity − based Random 0 0 10 20 30 40 50 60 70 Number of replications

User.Experience.Demo.

Summary. • Online$social$network$fundamentally$changes$video$distribuHon$ landscape$ • PropagaHon9based$video$distribuHon$by$jointly$considering$user,$ content$and$context$ • A$data9driven$approach$unveils$the$propagaHon$paEerns$and$ predicHons$ ! General$propagaHon$paEerns$for$architecture$ ! Specific$propagaHon$predicHons$for$strategies$ • 294$Hmes$improvement$for$local$video$download$

Thank you! More$InformaHon:)$ hEp://www.zhiwang.info$