Analysis of the Spread of Information on Twitter
Description: Analysis of the Spread of Information on Twitter and its Application to Internet Content Distribution Department of Computer Science and Engineering, IIT Delhi By Amit Ruhela 2007CSZ8359 Motivation 2 Rapid growth in usage of OSN websites by
Related Topics
Download Presentation
"Analysis of the Spread of Information on Twitter" is the property of its rightful owner. Permission is granted to download and print the materials on this website for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
Presentation Transcript
slide1. Analysis of the Spread of Information on Twitter and its Application to Internet Content Distribution Department of Computer Science and Engineering,
IIT Delhi By
Amit Ruhela
2007CSZ8359<br>
slide2. Motivation 2 Rapid growth in usage of OSN websites by individuals, celebrities, businesses and advertisers
31% of traffic directed on more than 300,000+ popular websites is produced by 400M users of top 8 OSN websites
( Shareholic Statistics )
ISP/CDN providers who serve a substantial portion of Internet traffic, are looking for better solutions to manage and deliver web content<br>
slide3. 3 Outline<br>
slide4. Datasets 4 Details of Kaist Dataset : H. Kwak, C. Lee, H. Park, S. Moon, “What is Twitter, a social network or a news media?”, in WWW ’10, New York, USA, 2010 Details of Twitter7 Dataset : J. Yang, J. Leskovec. Temporal Variation in Online Media. In WSDM '11, 2011.,<br>
slide5. Geographical Analysis 5 Conclusions :
Popular topics cross regional boundaries while unpopular topics stay within them.
Geographic crossovers can indicate popularity growth to guide CDN caching<br>
slide6. Events Based Analysis 6 Conclusion :
Most users tweeting on popular topics form one large connected component while unpopular topics are discussed in disconnected clusters.
The giant component forms when many tightly clustered sets of users discussing a topic coalesce together.
Tracking change in component size of topics can detect virality in advance and can possibly guide CDN caches to pre-fetch popular content . Ratio of size of largest to 2nd largest component<br>
slide7. Conclusions Popular topics cross regional boundaries while unpopular topics stay within them.
Most of the people talking about a popular topic on a given day tend to form a large connected subgraph (giant component)
The giant component forms when many tightly clustered sets of users discussing the topic merge together. 7<br>
slide8. 8 Outline<br>
slide9. Spatial Pattern Insight :
Time-lag across different time-zones can possibly suggest the time at which content should be pre-fetched in CDN servers 9 Event Time
US 2:26 PM.
India 2:56 AM Related Work :
N. Sastry, E. Yoneki, and J. Crowcroft, “Buzztraq: predicting geographical access patterns of social cascades using social networks” in Proceedings of the Second ACM EuroSys Workshop on Social Network Systems. Germany: ACM, 2009,
“ Knowledge about the number and location of friends of previous users can be used to generate hints that enable placing replicas of the content closer to future accesses”<br>
slide10. Insight :
Temporal event detection algorithms can model growth, decay, stability and periodicity of topic popularity to guide CDN caching Temporal Pattern 10 Growth in Popularity
Decay in Popularity
Lifespan of Topics
Periodicity of Popularity<br>
slide11. Social Pattern Insight:
To guide caching of content related to long tailed topics, the immediate social network neighbourhood can indicate pockets of popularity 11 Social Cohesion = Actual count of social relations between UGC producers
( per topic) ------------------------------------------------------------------------------------------------
Maximum possible social relations between UGC producers Related Work :
S. Scellato, C. Mascolo, M. Musolesi, and J. Crowcroft “Track globally, deliver locally: improving content delivery networks by tracking geographic social cascades" In Proceedings of the 20th international conference on World wide web (WWW '11). ACM, New York, USA, 2011
Used geographic information extracted from social cascades to improve the caching of multimedia files in a Content Delivery Network.<br>
slide12. Time-zone information can be used to determine when content would be requested in other geographies.
Temporal event detection algorithms can model growth and decay phases to guide CDN caching
To guide caching of content related to long tailed topics, the immediate social network neighbourhood can indicate pockets of popularity. 12 To design efficient content placement strategies: Conclusions<br>
slide13. 13 Outline<br>
slide14. Practically infeasible to track all OSN users and all the trending topics
Can the tracking of trends on OSNs be optimized by following just few users? Motivation 14<br>
slide15. Datasets 15 * Captures 60% of the entire Twitter user base from India<br>
slide16. Users Classification 16 Related Work :
M. Cha, F. Benevenuto, H. Haddadi and K. Gummadi, "The World of connections and Information Flow in Twitter," in IEEE Transactions on Systems, Man, and Cybernetics - Part A: Systems and Humans, 2012.
Classified users among Popular, Evangelist and Grassroots on basis of CDF of incoming links.<br>
slide17. 17 Tweeting Volume and Adoption Observation :
The volume of tweets by popular vs ordinary users is not distinguishable from one another.
Popular users start tweeting sooner than ordinary users by approximately 10% of the growth phase duration.<br>
slide18. 18 Participation of both popular and ordinary users decays over time Observation : Popular users participate longer in an event.<br>
slide19. 19 Content copying characteristics Observation :
Popular users write more original tweets than retweets by a factor of 60:40.
The tweets by popular users are retweeted 6 times more than tweets by other users. Time-Delay characteristics of content copying Observation :
Popular users are the quickest to retweet tweets by a factor of 8 than other users.
Popular users show a preference to retweet tweets by less popular users sooner, probably those who are their friends.<br>
slide20. 20 Influence on growth rates Observation :
Popular users do not seem to have influence on growth rate of events.
Level of popular users’ participation doesn’t say much on viraity of topics. Related Work :
Harrigan et al. in “Influentials, novelty, and social contagion: The viral power of average friends, close communities, and old news” : in Elsevier Social Networks, 34(4), 2012
Community structure rather than hubs that substantially increase social contagion on Twitter
Weng et al. in "Virality prediction and community structure in social networks" in Scientific Reports, 3(2522). 2013
Role of network structure is more likely to be a powerful driver for the emergence of trends
Sharad Goyal et al. in “The Structural Virality of Online Diffusion” in Management Science 62 (1), 2015
It’s not only the viral spreading by which a piece of content spread but also mass media or marketing efforts that rely on “broadcast” mechanism
Domingos and Richardson in "Mining the network value of customers." in KDD ‘01
Key factors in determining influence are the relationship among ordinary users and the readiness of the social network to accept a novel item.<br>
slide21. Conclusions 21 The indicators based on popular users are slightly leading over the bulk of the population but they don’t give a lot more or earlier information than ordinary users.
Popular users do not predict which events will become popular, hence at best tracking popular users can give a reflection of how the overall population will behave, and hence can be used as markers instead of tracking all the users Related Work :
G. arc´ıa Herranz, E. M. Egido, M. Cebri´an, N. A. Christakis, J. H. Fowler, “Using friends as sensors to detect global-scale contagious outbreaks”, PLoS ONE abs/1211.6512 (4) (2014)
Information collected from a few randomly selected individuals and their friends, can detect contagious disease outbreaks in advance.
Muhammad Bilal Zafar, Parantapa Bhattacharya, Niloy Ganguly, Saptarshi Ghosh, and Krishna P. Gummadi. 2016. On the Wisdom of Experts vs. Crowds: Discovering Trustworthy Topical News in Microblogs. In Proceedings of the 19th ACM Conference on Computer-Supported Cooperative Work & Social Computing (CSCW '16). ACM, New York, NY, USA
For discovering news-stories related to a topic, it is sufficient to analyze only the tweets posted by a small number of experts on the topic than the global Twitter population(Crowd).<br>
slide22. 22 Outline<br>
slide23. Motivation 23 Limitation of CDNs :
Limited content can be cached due to space constraints
Problem :
How to improve cache replacement strategies in CDNs? Prior Work
Use OSN signals to guide content placement strategies :
OSN websites know which topics are trending
Long tail content can impede LRU performance Long-tail of content Related Work :
P. Gill, M. Arlitt, Z. Li, and A. Mahanti. “YouTube Traffic Characterization: A View from the Edge”. In IMC), San Deigo, CA, 2007.
Video requests follow a Zipf-like distribution
M. Cha, H. Kwak, P. Rodriguez, Y. Ahn, and S. Moon. “Analyzing the video popularity characteristics of Large-Scale user generated content systems”, IEEE/ACM Transactions on Networking 17(5), 2009.
Video popularity follows a power-law distribution with an exponential cutoff.
S. Scellato et.al. "Track globally, deliver locally: improving content delivery networks by tracking geographic social cascades" In Proceedings of the 20th international conference on World wide web (WWW '11). ACM, New York, USA, 2011
Used geographic information extracted from social cascades to improve the caching of multimedia files in a Content Delivery Network.<br>
slide24. Performance of LRU vs OSN aided Event-based caching 24 Observation :
OSN aided event-based algorithm perform better only when the cache size is small.<br>
slide25. Residence Time of Content Objects in the Cache 25 Observation :
Content objects of popular topics may never get evicted from the cache. 25<br>
slide26. Removal of ‘K’ Most Popular Topics 26<br>
slide27. Workload Variants Changing number of surrogate servers
Changing count of objects per topic
Changing content selection behavior 27 Observation :
Consistent trend of naïve LRU outperforming OSN aided caching algorithms at larger cache sizes.<br>
slide28. Supplementary Analysis: Cache Hits Vs Popularity Distribution 28 Observation :
LRU performs very well as the popularity distribution becomes more skewed. Region of Study Related Work :
P. R. Jelenkovi´c, A. Radovanovi´c, and M. S. Squillante. “Critical sizing of LRU caches with dependent requests” Journal of Applied Probability, 43:1013–1027, 2006.
LRU performs as well on a dependent sequence of random requests as it does on an independent sequence of requests with the same request frequencies, as long as the cache size is large enough.<br>
slide29. Conclusions 29 Assuming OSN-based workloads to be representative of the web workload, the wide range of parameters over which LRU outperforms OSN-aided event-based caching algorithms indicates that the trend should hold in general.
People are interested in general global content and not just local community centric content. This results in a Zipfian like popularity distribution in web workload, and therefore even long tailed content are not able to disrupt LRU performance.<br>
slide30. Conclusions from Research Work Tracking of OSN trends can be optimized by marking popular users only.
Popular users don’t seem to have influence on the growth rate of events.
OSN-aided cache replacement algorithms have limited scope for CDNs handling Web workloads. 30<br>
slide31. Thanks for listening 31 @ruhela_amit<br>
IIT Delhi By
Amit Ruhela
2007CSZ8359<br>
slide2. Motivation 2 Rapid growth in usage of OSN websites by individuals, celebrities, businesses and advertisers
31% of traffic directed on more than 300,000+ popular websites is produced by 400M users of top 8 OSN websites
( Shareholic Statistics )
ISP/CDN providers who serve a substantial portion of Internet traffic, are looking for better solutions to manage and deliver web content<br>
slide3. 3 Outline<br>
slide4. Datasets 4 Details of Kaist Dataset : H. Kwak, C. Lee, H. Park, S. Moon, “What is Twitter, a social network or a news media?”, in WWW ’10, New York, USA, 2010 Details of Twitter7 Dataset : J. Yang, J. Leskovec. Temporal Variation in Online Media. In WSDM '11, 2011.,<br>
slide5. Geographical Analysis 5 Conclusions :
Popular topics cross regional boundaries while unpopular topics stay within them.
Geographic crossovers can indicate popularity growth to guide CDN caching<br>
slide6. Events Based Analysis 6 Conclusion :
Most users tweeting on popular topics form one large connected component while unpopular topics are discussed in disconnected clusters.
The giant component forms when many tightly clustered sets of users discussing a topic coalesce together.
Tracking change in component size of topics can detect virality in advance and can possibly guide CDN caches to pre-fetch popular content . Ratio of size of largest to 2nd largest component<br>
slide7. Conclusions Popular topics cross regional boundaries while unpopular topics stay within them.
Most of the people talking about a popular topic on a given day tend to form a large connected subgraph (giant component)
The giant component forms when many tightly clustered sets of users discussing the topic merge together. 7<br>
slide8. 8 Outline<br>
slide9. Spatial Pattern Insight :
Time-lag across different time-zones can possibly suggest the time at which content should be pre-fetched in CDN servers 9 Event Time
US 2:26 PM.
India 2:56 AM Related Work :
N. Sastry, E. Yoneki, and J. Crowcroft, “Buzztraq: predicting geographical access patterns of social cascades using social networks” in Proceedings of the Second ACM EuroSys Workshop on Social Network Systems. Germany: ACM, 2009,
“ Knowledge about the number and location of friends of previous users can be used to generate hints that enable placing replicas of the content closer to future accesses”<br>
slide10. Insight :
Temporal event detection algorithms can model growth, decay, stability and periodicity of topic popularity to guide CDN caching Temporal Pattern 10 Growth in Popularity
Decay in Popularity
Lifespan of Topics
Periodicity of Popularity<br>
slide11. Social Pattern Insight:
To guide caching of content related to long tailed topics, the immediate social network neighbourhood can indicate pockets of popularity 11 Social Cohesion = Actual count of social relations between UGC producers
( per topic) ------------------------------------------------------------------------------------------------
Maximum possible social relations between UGC producers Related Work :
S. Scellato, C. Mascolo, M. Musolesi, and J. Crowcroft “Track globally, deliver locally: improving content delivery networks by tracking geographic social cascades" In Proceedings of the 20th international conference on World wide web (WWW '11). ACM, New York, USA, 2011
Used geographic information extracted from social cascades to improve the caching of multimedia files in a Content Delivery Network.<br>
slide12. Time-zone information can be used to determine when content would be requested in other geographies.
Temporal event detection algorithms can model growth and decay phases to guide CDN caching
To guide caching of content related to long tailed topics, the immediate social network neighbourhood can indicate pockets of popularity. 12 To design efficient content placement strategies: Conclusions<br>
slide13. 13 Outline<br>
slide14. Practically infeasible to track all OSN users and all the trending topics
Can the tracking of trends on OSNs be optimized by following just few users? Motivation 14<br>
slide15. Datasets 15 * Captures 60% of the entire Twitter user base from India<br>
slide16. Users Classification 16 Related Work :
M. Cha, F. Benevenuto, H. Haddadi and K. Gummadi, "The World of connections and Information Flow in Twitter," in IEEE Transactions on Systems, Man, and Cybernetics - Part A: Systems and Humans, 2012.
Classified users among Popular, Evangelist and Grassroots on basis of CDF of incoming links.<br>
slide17. 17 Tweeting Volume and Adoption Observation :
The volume of tweets by popular vs ordinary users is not distinguishable from one another.
Popular users start tweeting sooner than ordinary users by approximately 10% of the growth phase duration.<br>
slide18. 18 Participation of both popular and ordinary users decays over time Observation : Popular users participate longer in an event.<br>
slide19. 19 Content copying characteristics Observation :
Popular users write more original tweets than retweets by a factor of 60:40.
The tweets by popular users are retweeted 6 times more than tweets by other users. Time-Delay characteristics of content copying Observation :
Popular users are the quickest to retweet tweets by a factor of 8 than other users.
Popular users show a preference to retweet tweets by less popular users sooner, probably those who are their friends.<br>
slide20. 20 Influence on growth rates Observation :
Popular users do not seem to have influence on growth rate of events.
Level of popular users’ participation doesn’t say much on viraity of topics. Related Work :
Harrigan et al. in “Influentials, novelty, and social contagion: The viral power of average friends, close communities, and old news” : in Elsevier Social Networks, 34(4), 2012
Community structure rather than hubs that substantially increase social contagion on Twitter
Weng et al. in "Virality prediction and community structure in social networks" in Scientific Reports, 3(2522). 2013
Role of network structure is more likely to be a powerful driver for the emergence of trends
Sharad Goyal et al. in “The Structural Virality of Online Diffusion” in Management Science 62 (1), 2015
It’s not only the viral spreading by which a piece of content spread but also mass media or marketing efforts that rely on “broadcast” mechanism
Domingos and Richardson in "Mining the network value of customers." in KDD ‘01
Key factors in determining influence are the relationship among ordinary users and the readiness of the social network to accept a novel item.<br>
slide21. Conclusions 21 The indicators based on popular users are slightly leading over the bulk of the population but they don’t give a lot more or earlier information than ordinary users.
Popular users do not predict which events will become popular, hence at best tracking popular users can give a reflection of how the overall population will behave, and hence can be used as markers instead of tracking all the users Related Work :
G. arc´ıa Herranz, E. M. Egido, M. Cebri´an, N. A. Christakis, J. H. Fowler, “Using friends as sensors to detect global-scale contagious outbreaks”, PLoS ONE abs/1211.6512 (4) (2014)
Information collected from a few randomly selected individuals and their friends, can detect contagious disease outbreaks in advance.
Muhammad Bilal Zafar, Parantapa Bhattacharya, Niloy Ganguly, Saptarshi Ghosh, and Krishna P. Gummadi. 2016. On the Wisdom of Experts vs. Crowds: Discovering Trustworthy Topical News in Microblogs. In Proceedings of the 19th ACM Conference on Computer-Supported Cooperative Work & Social Computing (CSCW '16). ACM, New York, NY, USA
For discovering news-stories related to a topic, it is sufficient to analyze only the tweets posted by a small number of experts on the topic than the global Twitter population(Crowd).<br>
slide22. 22 Outline<br>
slide23. Motivation 23 Limitation of CDNs :
Limited content can be cached due to space constraints
Problem :
How to improve cache replacement strategies in CDNs? Prior Work
Use OSN signals to guide content placement strategies :
OSN websites know which topics are trending
Long tail content can impede LRU performance Long-tail of content Related Work :
P. Gill, M. Arlitt, Z. Li, and A. Mahanti. “YouTube Traffic Characterization: A View from the Edge”. In IMC), San Deigo, CA, 2007.
Video requests follow a Zipf-like distribution
M. Cha, H. Kwak, P. Rodriguez, Y. Ahn, and S. Moon. “Analyzing the video popularity characteristics of Large-Scale user generated content systems”, IEEE/ACM Transactions on Networking 17(5), 2009.
Video popularity follows a power-law distribution with an exponential cutoff.
S. Scellato et.al. "Track globally, deliver locally: improving content delivery networks by tracking geographic social cascades" In Proceedings of the 20th international conference on World wide web (WWW '11). ACM, New York, USA, 2011
Used geographic information extracted from social cascades to improve the caching of multimedia files in a Content Delivery Network.<br>
slide24. Performance of LRU vs OSN aided Event-based caching 24 Observation :
OSN aided event-based algorithm perform better only when the cache size is small.<br>
slide25. Residence Time of Content Objects in the Cache 25 Observation :
Content objects of popular topics may never get evicted from the cache. 25<br>
slide26. Removal of ‘K’ Most Popular Topics 26<br>
slide27. Workload Variants Changing number of surrogate servers
Changing count of objects per topic
Changing content selection behavior 27 Observation :
Consistent trend of naïve LRU outperforming OSN aided caching algorithms at larger cache sizes.<br>
slide28. Supplementary Analysis: Cache Hits Vs Popularity Distribution 28 Observation :
LRU performs very well as the popularity distribution becomes more skewed. Region of Study Related Work :
P. R. Jelenkovi´c, A. Radovanovi´c, and M. S. Squillante. “Critical sizing of LRU caches with dependent requests” Journal of Applied Probability, 43:1013–1027, 2006.
LRU performs as well on a dependent sequence of random requests as it does on an independent sequence of requests with the same request frequencies, as long as the cache size is large enough.<br>
slide29. Conclusions 29 Assuming OSN-based workloads to be representative of the web workload, the wide range of parameters over which LRU outperforms OSN-aided event-based caching algorithms indicates that the trend should hold in general.
People are interested in general global content and not just local community centric content. This results in a Zipfian like popularity distribution in web workload, and therefore even long tailed content are not able to disrupt LRU performance.<br>
slide30. Conclusions from Research Work Tracking of OSN trends can be optimized by marking popular users only.
Popular users don’t seem to have influence on the growth rate of events.
OSN-aided cache replacement algorithms have limited scope for CDNs handling Web workloads. 30<br>
slide31. Thanks for listening 31 @ruhela_amit<br>