1. Opinion Dynamics
Modelling
Social Impact Theory and the 2011 Portuguese
Elections
A. Fonseca, J. Louçã - ECCS 2011
2. Summary:
• Online social network data on 2011 portuguese elections
• Social impact theory
• Multi-agent based modelling
• Data validation
• Conclusions
A. Fonseca, J. Louçã - ECCS 2011
3. Online data collection:
• Data collected between the 30th October 2010 and the 21th of January 2011 and
between 27 March and the 6th of June 2011.
• Community of 1903 Twitter users after data cleaning.
• Set of 44 news feeds from media on Twitter (TV, Journals, Radio).
• Data filtered though a set of keywords.
• Average of circa 200 daily tweets.
• Complete tweets collected. In this study we use:
• daily_quantity_of_tweets(candidate)
• daily_quantity_of_tweets(party)
• Analysis of expression magnitude instead of expression content.
• Data available at http://www.theobservatorium.eu/elections.
A. Fonseca, J. Louçã - ECCS 2011
4. Presidential elections:
• Six Candidates
• Final Results:
Candidate Result
Cavaco 53,14%
Alegre 19,67%
Nobre 14,04%
Lopes 7,05%
Moura 4,52%
Coelho 1,58%
A. Fonseca, J. Louçã - ECCS 2011
11. Legislative elections:
• Five major parties
• Final Results*:
Party Result
PSD 41,19%
PS 30,42%
CDS/PP 12,72%
CDU 8,61%
BE 5,69%
*After normalization between major parties
A. Fonseca, J. Louçã - ECCS 2011
18. Experimental data conclusions:
• Users tend to tweet proportionaly to the quantity of news.
• Users tend to tweet about the news of the same day.
• The relative magnitude of tweeting between candidates/parties is similar to the results
obtained from classical pools (telephone, presential), with less accuracy however.
• The final election results can roughly be estimated by the magnitude of twitting either
from common users or the news media. [Véronis J., 2007] [Tumasjan A. et al, 2010]
A. Fonseca, J. Louçã - ECCS 2011
21. Model of political debate:
• Media influences agents, agents influence each other.
• Each agent has propensity or aversion ( ) for expressing about certain candidate/
party of +1 or -1 respectively.
• Each agent has a potential of supportiveness and persuasiveness that is the same for
each of its neighbors and that have a Normal probability distribution over the set of
agents.
• Agent i talks about A if the impact on agent i about A is above average in relation to
all the other impacts (about B, C, ...).
• The news media acts as an autonomous agent over a fraction of agents on the
community with zero supportiveness and equal persuasiveness towards its audience.
A. Fonseca, J. Louçã - ECCS 2011
22. Simulation - 1 run legislatives
A. Fonseca, J. Louçã - ECCS 2011
24. Experimental validation:
• Agent community as real community (1903 users, 46423 links).
• Input stimulus, media twitting , is given to 10% of agents (~190 agents).
• Media twitting is processed as normal inter-agent stimulus processing.
• Average of 20 runs.
• Benchmark:
‣ The cosine similarity between MABS ‘twitting’ and real community tweets.
• Variants:
‣ Network topology (random link rewiring and lattice network)
‣ Media coverage (percentage of media receptors)
‣ Lagged impact (lagged positive or negative aditional social impact)
A. Fonseca, J. Louçã - ECCS 2011
28. ‣ Better similarity (lower degree between multidimentional vectors) on lower
randomizations and on lattice..
‣ There seems to be a worst case at randomization 50% but overall there is no significant
dependency on topology.
‣ Lattice network has a good performance on news reproduction as agents discuss with
lesser agents (avrg degree ~ 4).
‣ Network ‘hub’ tends to impose its opinion over the community [Atay F., 2006].
A. Fonseca, J. Louçã - ECCS 2011
29. ‣ Greater media coverages increase cosine similarity.
‣ Greater dependency on the lattice network.
A. Fonseca, J. Louçã - ECCS 2011
30. ‣ Large influence at 100% (k = 1) delayed impact in Legislatives. Day
after debate? Need validation from content analysis.
‣ Uncharacteristic at Presidentials.
A. Fonseca, J. Louçã - ECCS 2011
31. Multi-agent model conclusions:
• Cosine similarity between agent expression and real community is high.
• Expression is subject to an internal ‘subjective’ election replicated at community scale.
• Simulation similarity with real tweets not dependent on expression magnitude.
• Good difusion of media information.
• Non uniform influence of delayed impact of discussion.
• Dependence on the network topology.
‣ Hub nodes tend to influence community debates.
‣ Debate on ‘lattice’ tend to replicate know information from media.
‣ Scale-free seems to be most favorable topology for debate influence on overall
community.
A. Fonseca, J. Louçã - ECCS 2011
32. Future Work:
• To better qualify topology dependance.
• The role of Sij and Pij parameters.
• Examine other stimulus selection mechanism other than ‘greather than average’.
• Content analysis.
A. Fonseca, J. Louçã - ECCS 2011
33. Some references:
[Latané, B, 1981] Latané, B. (1981). The psychology of social impact, American Psychologist 36, 343-356.
[Atay, F. 2006] Atay, F. M., T. Biyikoglu, and J. Jost, Synchronization of networks with prescribed degree, IEEE Trans.
Circuits Syst. I 53(1):92–98 (2006).
[Tumasjan et al, 2010] Tumasjan, A., Sprenger, T. O., Sandner, P. G., and Welpe, I. M. (2010). Predicting
Elections with Twitter : What 140 Characters Reveal about Political Sentiment. In Word Journal Of The
International Linguistic Association, pages 178–185.
[Véronis, 2007] Véronis, J. (2007).Citations dans la presse et résultats du premier tour de la présidentielle
2007. Technical report.
Castellano, C., Fortunato, S, and Loreto, V. (2007). Statistical physics of social dynamics. Reviews of Modern
Physics, pages 1-58.
Connor, B. O. Balasubramanyam, R., Routledge, B. R. And Smitth, N.A. (2010) From Tweets to Polls: Linking
Text Sentiment to Public Opinion Time Series. In 4th Internation AAAI Conference on Weblogs and Social
Media, number May.
Gayo-avello, D. (2011). Limits of Electoral Predictions using Twitter. In ICSWSM-11 Barcelona, Spain.
Ghosh, R. And Lerman, K. (2010). Predicting Influential Users in Online Social Networks. In 4th SNA-KDD
Workshop at 16th ACM SIGKDD.
Holyst, J. A., Kacpersky, K., and Scheitzer, F. (2001). Social impact models of opinion dynamics. Annual Reviews
of Computational Physics, 48(22):253-273.
A. Fonseca, J. Louçã - ECCS 2011