Download presentation
Presentation is loading. Please wait.
Published byMaurice Gray Modified over 9 years ago
1
Computational modeling and big data Presentation for discussion forum “Theory and mechanisms of social interaction in the big data era” Son Bernadinet, May 6-8, 2013 Andreas Flache University of Groningen, Department of Sociology – ICS
2
Flache, ThefoDa discussion meeting– San Bernardino, May 2013 The “big data challenge” Strong version: “ Correlation supersedes causation, and science can advance even without coherent models, unified theories, or really any mechanistic explanation at all. ” Anderson, 2008 | 2
3
Flache, ThefoDa discussion meeting– San Bernardino, May 2013 A weaker version Having “more data” we need “less theory” to understand social processes. Example: predict performance of large organizations from structure of networks between employees (e.g. “small world”) without having to know the “content” of the actual network relations. | 3
4
Flache, ThefoDa discussion meeting– San Bernardino, May 2013 Why Polarization in political opinions? (DiMaggio et al. 1996, Evans 2003, Fischer et al 2009) 4 increasing variance of opinion distribution in some issues in some historical periods (e.g. Sexual morality; attitudes towards poor people; lifestyles; consumption tastes) Research on small groups: evidence of opinion differences increasing over time (van Knippenberg et al. 2007, Early et al. 2000, Feldman 1969) Political polarization in the internet (Lazer ea 2009)
5
Flache, ThefoDa discussion meeting– San Bernardino, May 2013 Why social diversity and polarization? A puzzle? Two powerful and general mechanisms in interpersonal interaction Homophily the more similar people are, the more they influence each other. Influence the more people influence each other, the more similar they become. How can there be stable diversity in opinions (e.g. on cultural issues) in a world where nobody is entirely disconnected from influence? 5
6
Flache, ThefoDa discussion meeting– San Bernardino, May 2013 | 6 A macro-social condition: complex networks ›Small world networks: local clustering and global integration: does this breed consensus? ›We argue that this depends on the micro-processes Flache, A., & Macy, M.W. 2011. Small Worlds and Cultural Polarization. Journal of Mathematical Sociology 35.1. Pp. 146-176.
7
Flache, ThefoDa discussion meeting– San Bernardino, May 2013 Microtheory 1: only homophily and ‘positive’ influence ›Following earlier social influence models (French etc) Positive influence: Move towards weighted average opinion (o) of local neighbours opinion 1 0 social influence
8
Flache, ThefoDa discussion meeting– San Bernardino, May 2013 ›Following earlier social influence models (French etc) Positive influence: Move towards weighted average opinion (s) of local neighbours Homophily: change relational weight w The more similarity opinions i-j, the more influence does j have Microtheory 1: only homophily and ‘positive’ influence
9
Flache, ThefoDa discussion meeting– San Bernardino, May 2013 Microtheory 2: xenophobia and ‘negative’ influence added (c.f. Macy ea, 2003; Baldassari & Bearman, 2007; Fent, Groeber & Schweitzer, 2007;…) Positive and negative influence local neighbours “pull” or “push” depending on weight w ij Homophily and xenophobia: change relational weight w average distance > zero positive weight, else negative distance + _ weight
10
Flache, ThefoDa discussion meeting– San Bernardino, May 2013 A computational experiment ›Add small proportion of “random ties” between initially disconnected clusters Compare effects micro-theory 1, and micro-theory 2 N=100, 20 “caves”, K=2, opinion initially uniform random | 10 p add tie = 0.003 Density D = 0.04 Clustering C = 1.0 Density D = 0.045 Clustering C = 0.814 Average path length L = ∞ Average path length L = 6.29
11
Flache, ThefoDa discussion meeting– San Bernardino, May 2013 Microtheory 1: homophily and positive influence | 11 Outcome polarization measure P P=0: consensus P=1 Perfect polarization in two opposing factions Adding random ties reduces polarization
12
Flache, ThefoDa discussion meeting– San Bernardino, May 2013 Microtheory 2: …plus xenophobia and negative influence | 12 Outcome polarization measure P P=0: consensus P=1 Perfect polarization in two opposing factions Adding random ties increases polarization
13
Flache, ThefoDa discussion meeting– San Bernardino, May 2013 Tentative conclusion “Small worlds and polarization” ›Knowing the structure is not enough ›Prediction of theories of social influence and selection depends decisively on microprocess: Only homophily and positive influence: Smaller world fosters consensus also xenophobia and negative influence: smaller world can increase polarization
14
Flache, ThefoDa discussion meeting– San Bernardino, May 2013 Big data and computational modelling ›Even if we had “big data” on the structure of social networks between all members of a large population (e.g. mobile phone calls, c.f. Eagle, Macy 2010, Science) ›… we need to know how “process sensitive” outcomes are in a given structure (computational modelling) ›… we need more and better data to know how people’s opinions affect each other in social interaction. “big data”? “high quality small data”: lab experiments 0nline experiments (but little control on subjects) “sentiment analysis” of online interactions? (not yet sufficiently accurate) | 14
15
Flache, ThefoDa discussion meeting– San Bernardino, May 2013 15 We conducted a series of 4 experiments with in total 443 subjects. Overall design: Measure subjects’ opinions on pre-selected issues. E.g. “0..100 percent of immigrants who come to the Netherlands for economic reasons should receive a residence permit. ” ›Pair subjects varying distance on opinions and other characteristics. ›Repeated sequence of exposure to others’ opinions, (exchange messages to influence each other) adjust opinions. ›Attractions (“weights”) were also measured repeatedly ›In some conditions, we manipulated initial attraction E.g. dictator games, football support, different moral positions Empirical research: experiments (Takács e.a. in progress)
16
Flache, ThefoDa discussion meeting– San Bernardino, May 2013 Results experiments (1) No effect initial opinion distance on direction of influence | 16 Negative shift positive shift
17
Flache, ThefoDa discussion meeting– San Bernardino, May 2013 Results experiments (2) Also no effect of (induced) initial disliking No evidence for negative influence in lab. | 17
18
Flache, ThefoDa discussion meeting– San Bernardino, May 2013 Polarization without negative influence A model based on persuasive argument theory (Mäs, Takács, Flache & Jehn, 2012, Organization Science) ›Agents differ by demographic characteristics and by opinion ›Opinion is constituted by number of pro- and con arguments held by agents (e.g. arg_vector ++---- opinion = -0.33 ) ›Homophily: the more similar agents, the more likely they interact ›Influence: if agent i interacts with j, then i adopts one of the arguments currently held by j and ‘forgets’ another argument previously held. | 18
19
Flache, ThefoDa discussion meeting– San Bernardino, May 2013 Polarization with maximal faultline strength | 19 But this only happens when initial opinion is sufficiently correlated with demographic attributes (demographically biased opinions)
20
Flache, ThefoDa discussion meeting– San Bernardino, May 2013 How “criss-crossing” agents prevent polarization A “criss-crossing” agent shares at least one demographic attribute with each of the subgroups a small probability of interaction between subgroups always remains The faultline can never be maximal Sooner or later arguments will be communicated between opposing subgroups and the system moves into consensus eventually | 20
21
Flache, ThefoDa discussion meeting– San Bernardino, May 2013 Strong faultline with three ‘criss-crossing’ agents (N=20) | 21 Polarization in the short run Consensus in the long run
22
Flache, ThefoDa discussion meeting– San Bernardino, May 2013 A bold generalization “Big data” on opinion formation are (only?) useful in combination with › (computational) modelling of underlying mechanisms and ›“high quality small data” | 22
23
Flache, ThefoDa discussion meeting– San Bernardino, May 2013 What if we had the ideal “big data”? ›Record of all interactions between individual members of a population (who does when interact with whom) ›Proxy: geographical residence of all members of the population, other places where they meet (employers, voluntary associations, …) ›Their opinions on a range of issues pre- post interaction (or in temporally fine-grained panel) ›A range of individual characteristics (demographic…) | 23
Similar presentations
© 2025 SlidePlayer.com. Inc.
All rights reserved.