National Academies Press: OpenBook

Dynamic Social Network Modeling and Analysis: Workshop Summary and Papers (2003)

Chapter: Polarization in Dynamic Networks: A Hopfield Model of Emergent Structure

« Previous: Accounting for Degree Distribution in Empirical Analysis of Network Dynamics
Suggested Citation:"Polarization in Dynamic Networks: A Hopfield Model of Emergent Structure." National Research Council. 2003. Dynamic Social Network Modeling and Analysis: Workshop Summary and Papers. Washington, DC: The National Academies Press. doi: 10.17226/10735.
×
Page 162
Suggested Citation:"Polarization in Dynamic Networks: A Hopfield Model of Emergent Structure." National Research Council. 2003. Dynamic Social Network Modeling and Analysis: Workshop Summary and Papers. Washington, DC: The National Academies Press. doi: 10.17226/10735.
×
Page 163
Suggested Citation:"Polarization in Dynamic Networks: A Hopfield Model of Emergent Structure." National Research Council. 2003. Dynamic Social Network Modeling and Analysis: Workshop Summary and Papers. Washington, DC: The National Academies Press. doi: 10.17226/10735.
×
Page 164
Suggested Citation:"Polarization in Dynamic Networks: A Hopfield Model of Emergent Structure." National Research Council. 2003. Dynamic Social Network Modeling and Analysis: Workshop Summary and Papers. Washington, DC: The National Academies Press. doi: 10.17226/10735.
×
Page 165
Suggested Citation:"Polarization in Dynamic Networks: A Hopfield Model of Emergent Structure." National Research Council. 2003. Dynamic Social Network Modeling and Analysis: Workshop Summary and Papers. Washington, DC: The National Academies Press. doi: 10.17226/10735.
×
Page 166
Suggested Citation:"Polarization in Dynamic Networks: A Hopfield Model of Emergent Structure." National Research Council. 2003. Dynamic Social Network Modeling and Analysis: Workshop Summary and Papers. Washington, DC: The National Academies Press. doi: 10.17226/10735.
×
Page 167
Suggested Citation:"Polarization in Dynamic Networks: A Hopfield Model of Emergent Structure." National Research Council. 2003. Dynamic Social Network Modeling and Analysis: Workshop Summary and Papers. Washington, DC: The National Academies Press. doi: 10.17226/10735.
×
Page 168
Suggested Citation:"Polarization in Dynamic Networks: A Hopfield Model of Emergent Structure." National Research Council. 2003. Dynamic Social Network Modeling and Analysis: Workshop Summary and Papers. Washington, DC: The National Academies Press. doi: 10.17226/10735.
×
Page 169
Suggested Citation:"Polarization in Dynamic Networks: A Hopfield Model of Emergent Structure." National Research Council. 2003. Dynamic Social Network Modeling and Analysis: Workshop Summary and Papers. Washington, DC: The National Academies Press. doi: 10.17226/10735.
×
Page 170
Suggested Citation:"Polarization in Dynamic Networks: A Hopfield Model of Emergent Structure." National Research Council. 2003. Dynamic Social Network Modeling and Analysis: Workshop Summary and Papers. Washington, DC: The National Academies Press. doi: 10.17226/10735.
×
Page 171
Suggested Citation:"Polarization in Dynamic Networks: A Hopfield Model of Emergent Structure." National Research Council. 2003. Dynamic Social Network Modeling and Analysis: Workshop Summary and Papers. Washington, DC: The National Academies Press. doi: 10.17226/10735.
×
Page 172
Suggested Citation:"Polarization in Dynamic Networks: A Hopfield Model of Emergent Structure." National Research Council. 2003. Dynamic Social Network Modeling and Analysis: Workshop Summary and Papers. Washington, DC: The National Academies Press. doi: 10.17226/10735.
×
Page 173

Below is the uncorrected machine-read text of this chapter, intended to provide our own search engines and external engines with highly rich, chapter-representative searchable text of each book. Because it is UNCORRECTED material, please consider the following text as a useful but insufficient proxy for the authoritative book pages.

162 Polarization in Dynamic Networks: A Hopfield Mode! of Emergent Structure Michael W. Macy, Comell University; James A. Kitts, University of Washington; Andreas Flache, University of Gron~ngen; Steve Benard, Cornell University ABSTRACT Why do populations often self-organize into antagonistic groups even In the absence of competition over scarce resources? Is there a tendency to demarcate groups of"us" and '`them" that is inscribed In our cognitive architecture? We look for answers by exploring the dynamics of influence and attraction between computational agents. Our mode} is an extension of Hopfield's attractor network. Agents are attracted to others with similar states (the principle of homophiTy) and are also influenced by others, as conditioned by the strength and valence of the social tie. Negative valence implies xenophobia (instead of homophily) and differentiation (instead of imitation). Consistent with earlier work on structural balance, we find that networks can self- organ~e into two antagonistic factions, without the knowledge or intent ofthe agents. We mode! this tendency as a Unction of network size, the number of potentially contentious issues, and agents' openness and flexibility toward alternative positions. Although we find that polarization into two antagonistic groups is a unique global attractor, we investigate the conditions under which uniform and pluralistic alignments may also be equilibria. From a random start, agents can self-organ~ze into pluralistic arrangements if the population size is large relative to the size of the state space. DYNAMIC SOCIAL NETWORK MODELING ED CYSTS

2 INTRODUCTION: GROUP POLARIZATION Why do populations often self-organize into antagonistic groups even in the absence of competition over scarce resources? Is there a tendency to demarcate groups of"us" and '`them" that is inscribed in our cognitive architecture? We look for answers by exploring the dynamics of influence and attraction between computational agents. Our model builds on recent work using agent-based models of cultural convergence and differentiation (1,2,3). Scholars contend that the abundance of social groups containing highly similar actors is due to dual processes of attraction and social influence. Formal models generally assume that each agent chooses interaction partners that are similar to itself, while social interaction also leads agents to adopt each other's traits and thus grow more similar. In this positive feedback loop, a minimal initial similarity increases the probability of interaction which then increases similarity. This process of consolidation (4) will presumably continue until all agents engaged in mutual interaction have converged to unanimity. While this self-reinforcing dynamic seems to imply a "melting pot," or inexorable march toward homogeneity, scholars have shown that global diversity can survive through impermeable barriers to interaction between distinct subcultures. If bridge ties between groups are entirely absent, then this local convergence can indeed lead to global differentiation, in spite of cultural conformity among the individual agents. Stable minority subcultures persist because of the protection of structural holes created by cultural differences that preclude interaction. Given such barriers, homogenizing tendencies actually reinforce rather than eliminate diversity. The models thus predict social stability only where a population is entirely uniform or where agents cluster into mutually exclusive uniform subpopulations that are oblivious to one another. This global pluriformity depends on the assumption that interaction across these boundaries is impossible. Any allowance for interaction between dissimilar agents, no matter how rare, leads diversity to collapse into global homogeneity. This also reflects earlier analytical results obtained with models of opinion dynamics in influence networks (21,22,23). Abelson proved that a very weak condition is sufficient to guarantee convergence of opinions to global uniformity: The network needs to be "compact," such that there are no subgroups that are entirely cut off from outside influences. This apparently ineluctable homogeneity in fully connected networks led Abelson (21, p. 153) to wonder "... what on earth one must assume in order to generate the bimodal outcome of community cleavage studies?" We will pursue this interest in cleavages, or structural bifurcations, as an alternative explanation for both the disproportionate homogeneity in social groups and the persistence of diversity across groups. First let us revisit the generative processes, the psychological and behavioral foundations of attraction and social influence. These models of convergence implement one of the starkest regularities in the social world: "homophily," or the tendency for each person to interact with similar others (5, 6, 7). Several explanations for homophily have been proposed. Social psychologists posit a "Law of Attraction" based on an affective bias toward similar others (8; see also 9,10,11,12). Even disregarding such an emotional bias, structural sociologists (3,2,14) point to the greater reliability and facility of communication between individuals who share vocabulary and syntax as an inducement for homophilous relations. Others counter that homophily is the spurious consequence of social influence, as in classic research on 'pressures to uniformity" (16,17,18) and recent work in social networks (19,20). Even if relations are held constant in an exogenously clustered social network, social DYNAMIC SOCIAL NETWORK MODES AND ISIS 163

3 In flu ence Tom network neighbors will lead to local homogeneity, without the need to assume homophilous interaction. Scholars have thus overwhelmingly attributed homophily In observed-social networks to some combination of differential attraction and social influence, where agents choose similar partners and partners grow more similar over time. Much less attention has been directed to an alternative explanation - that homophily is largely a byproduct of its antipole. It is not so much attraction to those who are similar that produces group homogeneity but repulsion Tom those who are different. Xenophobia leads to the same emergent outcome as attraction between similar actors: disproportionate homogeneity in relations. In fact, Rosenbaum (15) argues that many experimental flndiIlgs of homophily in relations may have spuriously represented this effect of repulsion Dom those who are different. The most sustained treatment of positive and negative ties has appeared in the literature on Balance Theory. Following Heider (24), scholars have assumed that actors are motivated to mamta~n "balance" in their relations, such that two actors who are positively tied to one another will fee! tension when they disagree about some third cognitive object. Formally, we can think of this as a valued graph of agents A and B along with object X, where the graph will be balanced only when the sign product A *Boxes positive. For example, if A and B are positively tied but A positively values X and B negatively values X, then the graph is imbalanced. In order for this dissonance to resolve and result In a stable alignment, there will either be a falling out between A and B or one (but not both) of these actors will switch evaluations of X A n~rnile1 Hen operates if the tie between A and B is negative. - r~~~ r^~-_~V A prolific line of research- Structural Balance Theory (32) - has examined a special case of Heider's model, where the object X is actually a third actor, C. The mode! simply extends to tnadic agreement among agents A, B. and C, where balance obtains when the sign product of the tnad A-B-C is positive. This formalizes the adage that a Fiend of a friend or an enemy of an enemy is a friend, while an enemy of a Fiend or a Fiend of an enemy is an enemy. Extended to a larger network, it is well known that the elemental triadic case suggests a perfect senar~tion of two mutually antagonistic subgroups. Our mode} integrates the attraction-~nfluence feedback loop from formal models of social convergence with the bivalent relations In Balance Theory. Following Nowak and ValIacher (29), the model is an application of Hopfield's attractor network (25, 26) to social networks. Like Heider's Balance Theory, an important property of attractor networks is that individual nodes seek to minimize "energy,' (or dissonance) across all relations with other nodes. As we will see, this suggests self-re~nforc~ng dynamics of attraction and influence as well as repulsion and differentiation. -look ~ rid row More precisely, this class of models generally uses complete networks, with each node characterized by one or more binary or continuous states and linked to other nodes through endogenous weights. Like other neural networks, attractor networks leant stable configurations by iteratively adjusting the weights between individual nodes, without any global coordination. In this case, the weights change over time through a Hebbian learning rule (27~: the weight ~v'; is a Unction of the correspondence of states for nodes i and j over time. To the extent that i and j tend to occupy the same states at the same time, the tie between them will be increasingly positive. To the extent that i and j occupy discrepant states, the tie will become increasingly negative. 164 DYNAMIC SOCIAL N~TWORKMODELI?JG AND ANALYSIS

4 Extending recent work (30), we apply the Hopfield model of dynamic attraction to the study of polarization in social networks. In this application, observed sunilarity/difference between states determines the strength and valence of the tie to a given referent. MODEL DESIGN In our application of the Hopfield model, each node has N-1 undirected ties to other nodes. These ties include weights, which determine the strength and valence of influence between agents. Formally, social pressure on agent i to adopt a binary state s (where s= +1) is the sum of the states of all other agents j, conditioned by the weight (wit) of the dyadic tie between i andj (-~.0 < wit ~ I.0): N ~ WijSj Pw=i=' , Jo (1) Thus, social pressure (-<A<) to adopt s becomes increasingly positive as ''s "friends" adopt s (so=) and i's "enemies" reject s (sari). The pressure can also become negative in the opposite circumstances. The model extends to multiple states In a straightforward way, where tI] independently determines the pressure on agent i for each binary state s. Strong positive or negative social pressure does not guarantee that an agent will accommodate, however. It is effective only if i is willing and able to respond to peer influence. If i is closed-minded or if a given trait Is not under i's control (e.g., ethnicity or gender), then no change to s will occur. Let v classify states as fixed (v=0) or Dee (my) and X represent an exogenous determination of s. We then find i's propensity Liz to adopt state s as a cumulative logistic Unction of social pressure: This = I, + A-,, + (1 vS)Xi (2) Agent ~ adopts s if ~z>0.5+%s, where % is a random number (-0.5<%<0.5) and ~ is an exogenous error parameter (0~<~. At one extreme, - 0 produces deterministic behavior, such that any social pressure above the trigger value always leads to s=l and pressure even slightly below the trigger value always leads to sew. FoHow~ng Harsanyi (31), s>0 allows for a "smoothed best reply" ~ which pressure levels near the trigger point leave the agent relatively indifferent and thus likely to explore behaviors on either side ofthe threshold. In the Hopfield model, the path weight wit changes as a function of similarity In the states of node i and j. Weights begin with uniformly distributed random values, subject to the constraints that weights are symmetric (wifwji). Across a vector of K distinct states so, (or the position of agent i on issue k), agent i compares its OWD states to the observed states of another agent j and adjusts the weight upward or downward corresponding to their aggregated level of agreement or disagreement. Based on the correspondence of states for agents i end j, their weight will change at each discrete hme point t ~ proportion to a parameter i, which defines the rate of structural learning (0<~<~: ~ K Wij,t+1 Wijt (1 2) +—~Sj~SiEt' j ~ i (3) K k=1 DYN~MICSOCIALN~ TWORKMODE~G~D~^YSIS 165

As correspondence of states can be positive (agreement) or negative (disagreement), ties can grow positive or negative over time, with weights between any two agents always symmetric. Note that both processes - adjustment of social ties and adjustment of behavior - seek to maximize balance ~ relations between any two agents across their vectors of states. Given an initially random configuration of states and weights, these agents will search for a profile that ~ninim~zes dissonance across their relations. METHODS We tested the dynamics of network polanzation by manipulating three agent-level behavioral rules:flexibili~y, broaci-mindedness, and open-mindedness: I. Rigid vs. flexible (a.k.a. "wafflmg"~: the extent to which the state adoption decision is stochastic near indifference, ranging Dom s=0 (agent i always picks the strictly preferred position even when almost indifferent) to ~l (i adopts state s with propensity ~..s). 2. Narrow- vs. broad-minded: the multiplexity of the state space, that is, the number of salient cross-cutt~ng dimensions that agents consider in evaluating each other. 3. Closed- vs. open-minded: the proportion ofthese salient dunensions that can be affected by social pressure, with a minimum of at least one free state. A fixed state can be a position on an issue about which an agent is strictly uncompromising. It can also be an attribute that Is difficult to change, such as ethnicity or gender. Thus, we consider a range of stylized character profiles for agents, ranging Dom "extremists" who focus on ascriptive differences and are rigidly narrow- and closed-minded, to more broad-minded "moderates" who tend to waffle on the issues, and are open to influence (and/or focused on behavior rather than ascriptive traits). We measure network polarization as the degree of segregation among mutually exclusive cohesive subgroups of agents, measured after all weights and states have converged to equilibrium. This measure is based on the graph theoretical ES-set (321. An LS-set is a subset of agents who have stronger ties to members within the subset than they have to members outside the subset. For the cohesive subgroups thus identified, we calculated their segregation in terms of the normalized difference between the average internal tie strength and the average tie strength between the subgroup and its complement. To obtain an overall polarization score, we generated all partitionings of the graph into mutually exclusive cohesive subgroups and took Tom these the average segregation of all subgroups In the most segregated partitioning. We allowed each experimental condition to repeat for 1000 iterations, which was sufficient for almost all conditions to arrive at an equilibrium Only equilibrium outcomes (converged solutions) were analyzed. For all the experiments, we set the structural learning rate (~) at 0.5, which means that agents consider current states equally with previous impressions in determ~n~g the tie weight for the next Arbitrary time ~nterval.t Each parameter combination was repeated 20 times, y~eld~g a total of 44,000 observations. More technically, for each of the subsets of an LS-set, average tie strength between members of We subset and its complement within the LS set is higher than the average tie strength to members outside We LS set. We first identified subgroups of agents that agree on those states that they are Bee to change. We then searched through coalitions of these subgroups to find maximal LS-sets in the graph defined by agreement on Dee states. Finally, we used these LS-sets to identify the maximally segregated partitioning of the network, based on the relational weights (which reflect agreement on all states ~ fixed and Eee). Qualitative patterns shown appear to be robust to a very broad range of assumptions about the learning rate, though extremely low learning rates make waiting for convergence impractical. 166 DYNAMIC SOCIAL NETWORK MODELING ED ISIS

6 RESULTS Experiment ~ tests the effects on network polarization of agent broad-mindedness, while holding flexibility and openness constant (~0 and v=~. With zero flexibility and no fixed states, network dynamics converged to equilibrium In every case. Network polarization is measured as the degree of segregation In the most segregated equilibrium partitioning of the weight graph into mutually exclusive cohesive subgroups. Maximum segregation (~.0) occurs in a perfectly bifurcated network, composed of exactly two internally cohesive and mutually antagonistic groups. Broad-mindedness is operational~zed as multiplexity, or the number of dimensions In the state space. Intuitively, we would expect polarization to be most likely in a world where agents focus single-mindedly on one highly salient issue, and indeed, Figure ~ confirms the intuition. What is surprising, however, is that network polarization can also be caused by too many salient issues. With 100 agents scattered over a state space larger than 2s, the dynamics quickly approach the bifurcation we would expect in a single-issue population. As N declines, the non- monotonic effect becomes less pronounced. For N>100, the U-shape function is little changed but the critical value of K increases, Tom 5 dimensions with N<100 to I] dimensions with N=300. Polarization N '~ °1 2 Multple~aty 0.8 0.4 -0.2 Figure I. Effect of network size (IO~N<IOO) and agent multiplexity (~K<IO) on network polarization (with s=0 and v=~. Experiment ~ assumes rigidly deterministic agents (~=0) who never explore alternative positions, even when they approach indifference. It also assumes that agents are willing and able to change ad positions in response to social influence (v=~), which would not be the case, for example, if race or ethnicity were a salient dimension of social differentiation. Experiment 2 relaxes these assumptions under conditions that malce polarization relatively unlikely, namely, that there are at least five dimensions of differentiation (see Figure I).+ Intuitively, we might Fewer than five dimensions also constrains the ability to manipulate the relative number of dimensions that agents can alter. DYNAMIC SOCIAL NETWORK HODEL~G ED ^¢YSIS 167

7 expect less network polarization in a population that is less rigid and more polarization ~ a population that is less open-minded. The surprising result is that the elect is quite the opposite. Figure 2 displays the polarizing effects of flexibility and open mindedness in greater detail. As in Experiment I, polarization is based on the relational partition that maximizes segregation ofthe weight graph among coalitions of distinct subgroups in the matrix of Bee states. Flexibility is simply the error level, ranging from s=0 (always pick the strictly preferred position) to £=i (adopt the position with propensity At. Openness is the proportion of K states that are affected by social pressure (or Ivs/K), where there must always be at least one bee state. Results were averaged over 20 replications of each treatment condition, in which N ranged Mom O to 100 and K ranged Tom 5 to ~ O (hence openness ranged Tom 0.2 to ~ .0~. ~ Polanzation l Flexibility ~.2 0.4 0.6 ,S>~ ~ ~ I. < ~~ ~ Ives l +~ s ~ ?.~ its a; ~~-~-~<,~t my' As- · ~ ~~ 7 ';. i "I'm ~ ~ ~ Y it. ~— a''.',',,,, \ ~ ~ ~ ~ ~ \ we; ,,~,~,i~,,,s~,,.,,.~~ ~~,&~ ~ y ~~ ~ A, ~^ 1 >~! ~> I,; ~ ~~ ~ ~t \ W~ 1 ~ ~ ~ ~ S <I ~ ~ ~ ~ ~ ~ ~ X X ~ \ ~ ~ ~ If\ \~! _ ~~\ \ / ~ ~ \ \ ~~¢,,,,{~,:,, ~ ~ Id/\ \ ~- 0.8 0.6 0.4 0.8 Openness 0.2 Figure 2. Effect of agent flexibility (0~£<1) and open-minde~ess (~::vslK) on network polarization (with K=[5. . . 10], N=[ 10. . . 100]~. The surprising result is that polarization sharply increases with flexibility across all levels of openness. A population of "hardl~ners" (who never reveal ambivalence by wafting on the issues) is generally less prone to become polarized. Less surprisingly, polarization also increases with openness to social influence, for the simple reason that agents cannot alter their fixed states as a way of reducing dissonance in their relations. Note that Figure 2 is based only on equilibrium outcomes. As flexibility increases, the rate of convergence declines, due to the destabilizing effects of stochasticity. If we include nonconverged outcomes (coded with zero polarization), the effect of flexibility goes Dom convex to concave when most states are fixed, and when all states are Dee, the effect becomes non-monotonic, with polarization declining at maximum flexibility. In other words, with maximum flexibility and openness, the network is less likely to develop stable factions, but if this should happen, the number of factions is almost certain to be two. DISCUSSION With binary agent states, the dynamics of homophilous influence and xenophobic differentiation create an energy landscape in which there is only one basin of attraction, polarization. Any configuration with more than two cohesive subgroups represents global tension for a reason that is readily apparent: A binary state precludes the ability to hold a position that 168 DYNAMIC SOCIAL NF7TWORI<MODE:~WG kD ISIS

8 differs from each oftwo opposing positions. As the number of dimensions increases, it is possible for many more than two Unique combinations to persist at equilibrium, but for each, there must always be some sunilarity with all other combinations except one. Surprisingly, as these combinations explode, they lead not to pluralism but to polarization. The mechanism is similar to that identified by Axeirod. The more opportunities for neighborhoods to be liniced via overlapping positions, the higher the probability they will find a way to coalesce. The key variable here is the density ofthe population distribution across the state space. When there are many agents but few dimensions, then, from a random start, every possible combination of states (corresponding to "ideologies" if the states are variable and to ~`identities'' if the states are fixed) will find multiple incumbents. The presence of identical neighbors (as well as dissimilar antipodes who serve as negative referents) creates strong pressures to remain loyal to a shared set of states. This resistance to change can then support a pluriform equilibrium. Increasing the number of dimensions expands the set of possible combinations, allowing groups to mobilize around a variety of configurations of states. Accordingly, polarization initially decImes with increasing multiplexity. However, if population density In the state space fats below a critical level (either due to low population or high multiplexity), it becomes impossible for every distinct vector of states to find an incumbent. The sparseness of the population relative to the number of unique positions ~ state space requires that some agents begin with neither allies (who overlap perfectly3 nor enemies (who overlap not at all). These agents lack sufficient pressure to stick to their guns and will thus be pulled into coalitions with agents who overlap partially with their positions. As the multiplexity ofthe state space increases, equilibria ncreas~gly depend on a delicate balance, where a distnbution of agents on ideologies and identities needs to be found that corresponds to the distances between positions, such that they are not so close to each other that they are pulled together and not so far apart that they are pushed toward the opposite pole. In this higher range, increasing multiplexity makes it more difficult for pluriform configurations to persist, such that collapse into a simple bifurcation becomes the most likely alternative. Flexibility promotes polarization by the same process. Network self-organization can become trapped in high-energy pluriform equilibria that are nevertheless a local minimum ofthe energy landscape, as depicted in Figure 3. issues can disturb these local solutions, allowing the system to continue searching for the global attractors. Thus, increased flexibility washes out the non-monotonic effect of multiplexity evident ~ Figure I. Flexibility leads to polarization even when agents attend to multiple dimensions of differentiation that would otherwise produce the cross-cutt~ng cleavages characteristic of a pluriform society. Agents who express ambivalence by "waffling" on the I' \ ~ DYNAMIC SOCIAL NETWORK MODELING AND ANALYSIS 169

Figure 3. How agent volatility allows the network to find a lower energy minimum- hence greater polarization. Homogeneity is also a stable equilibrium of the model, but this equilibrium cannot obtain sinless we preclude negative weights and negative influence (as In the earlier models of cultural convergence and differentiation discussed above). If we initialize the mode} at homogeneity, it will remain there indefinitely. However, from a random start, the system cannot find the s~ngle- group solution without falling into an inescapable trap of is-group/out-group hostility. Further, a sufficient exogenous perturbation will disturb the unitary equilibrium and yield a two-group solution, but the reverse shin is much more difficult to obtain and is extremely unlikely. To explore this interpretation further, we measured the energy level of equilibria by summing over the product of agreement and tie strength across all dyads in the weight graph. As expected, we found that networks with perfect polarization in two opposing camps had an energy level of zero, while energy levels increased with the number of different positions in the state space that co-existed at equilibrium. That analysis examined tension in dyads for the variety of converged configurations. A continuous index of graph balance (35), assessing balance in triads and cycles of any higher length, yielded the same conclusion. Equilibria with either one or two internally cohesive groups represent maximum balance, while convergence at any higher number of groups implies imbalances relations in the social network. CONCLUSIONS AND IMPLICATIONS This study has explored polarizing tendencies of a self-organizing network, using a Hopfield mode} of dynamic attraction. The project integrates Axeirod's positive feedback mode} of ~nfluence-~nteraction with a bivalent mode! of social relations and cognitive balance. The Hopfield mode! emulates well-known processes of homophily, xenophobia, and influence Tom positive and negative referents. Thus, we may account for emergent social phenomena with a rigorous and parsimonious model, mcludina onlY basic behavioral orincinles that have horn broadly supported in experimental work. _^ I,^ .,._,^~.,_ - i,.. ~~ ~ _ TV The results have interesting implications for the notion of"structural balance" In social ~ .. . . ~ . . relations. Although the agents in this model are clearly designed to maintain balance in their behaviors with both positive and negative referents, this assumption is not '`~red m" to the relations themselves, as it is In Structural Balance Theory. That is, two agents A and B fee] no direct need for consistency ~ their relations with a third agent C. ~deed, A has no knowledge of the B-C relationship and thus no ability to adjust the ~4-B or A-C relations so as to balance the triad. NotablY. the results show that triads for cycle of TV lengthy Hm tend tn horning hulas ~ V A ~ ~ ~ _~) ~~ —_~ ~~ V_~v~~—~ v . . ~ . . . . . over time as agents seek balance in Shelf dvadic relatinn~ However there in no relaxant-- in this ~ 3 model that they will achieve a globally optimal state In structural balance' and we also observe equilibrium outcomes where more than two subgroups persist indefinitely. This outcome cannot be reconciled with Structural Balance Theory, which predicts that system-level stability can only occur when the group either has become uniform or has nc~larizer1 into two intern~liv unhip and mutually antipathetic cliques. --~ r ~ _~,,,_-,, _ The model also has interesting implications for Social Identity Theory (36), which posits an ~n-group bias toward those who share a salient trait, prejudice against the out-group, and a tendency to ignore or change discrepant traits. The Hopfield model produces dynamic networks that self-organize into a similar pattern, but without a higher-order cognitive Damework of social categories. In fact, these agents are not even aware that they belong to "groups" at all. This 170 9 DYNAMIC SOCIAL NETWO=MODEL~G^D ISIS

10 demonstrates the possibility that ~n-group/out-group differentiation and antagonism are emergent properties of network self-organ~zation and are not inscribed In agents' cognitive architectures as assumed by identity theonsts. In effect, the model demonstrates a distributed representation of the formation of social identities, which are not reducible to local representations In the minds of m~ividual agents. Fmally, the mode! offers an important methodological demonstration. The results show how analytical methods focusing on equilibria may be misleading for prediction of opinion distributions that are likely to obtain. An analysis of possible equilibria clearly shows that more multiplexity and larger population size increase the number of possible pluriform equilibria, i.e. Opinion configurations in which for all agents and all issues the aggregated pressures to stick to one's opinion exceed the aggregated pressures to change. However, our computational model showed that despite a larger number of theoretically possible equilibria, these conditions decreased the number of equilibria that were actually reached by the opinion dynamics. The reason is that equilibrium analysis fails to talce into account the ~tnlc.tilr~l 1erlrnino Her ~ ~ ~ I__ ~~= r~ IF _ —~7 ~ ~ ~ t ~ Inrougn wmcn outcomes are selected. 1I1 structural reaming, equilibria become increasingly unlikely as the complexity of the coordination process increases. As a consequence, the computational model identifies polarization as the global attractor in a multiplex opinion space, an important substantive result that is overlooked by static equilibrium analysis. Previous theoretical work has emphasized the global stability of social homogeneity, where convergence to unanimity is an almost irresistible force In closely interacting populations. If it is possible for some ties to be negative, our model suggests instead that a social structure is most stable when the network self-organ~zes along a single dimension of differentiation which thus determines a clear "right" and '~wrong" choice on all behavioral dimensions. In comparison, social homogeneity is highly brittle. ACKNOWLEDGEMENT The research was supported by a grant to the first author Dom the U.S. National Science Foundation (SES-0079381~. The research ofthe third author was made possible by a fellowship of the Royal Netherlands Academy of Arts and Sciences. REFERENCES . Axeirod, Robert. ~ 997. "The Dissemination of Culture: A Mode! with Local Convergence and Global Polarization." Journal of Conflict Resolution 41: 2023-226. 2. Mark, N. 1998. "Beyond individual differences: Social differentiation Dom first Dr~ncinies." Am Soc Rev 63:309-330. 3 . Carley, K. 1 99 1 . "A theory of group stability." Am Soc Rev 56:33 1 -54. 4. Latane, B. 2000. "Pressures to uniformity and the evolution of cultural norms." Pp. 189-215 in Computational Modeling of Behavior in Organizations: The Third Scientific Discipline, edited by D. R. Ilgen and C. L. Hulin. 5. Cohen, J. M. ~977. "Sources of Peer Group Homogeneity." Soc En! 50:227-41. 6. Kandel, D.B. 1978. "Homophily, Selection, and Socialization in Adolescent Friendships." Am Jour Soc 84:427-36. 7. Verbrugge, L.M. ~ 977. "The structure of adult Eiendship choices." Social Forces 56:576-97. 8. Byrne, D. E. 1971. The Attraction Paradigm. New York: Academic Press. 9. Newcomb, T.M. ~ 961. The Acquaintance Process. New York: Holt, Rhmehart & Winston. O. Hoffman, I.R. & Maier, N.R.F. 1966. "An experimental reexamination of the similarity- attraction hypothesis." Jour of Pers & Soc Psy 3 :145- ] 52. - r- --rip DYNAMIC SOCIAL NETWORK MODEM ED ISIS 171

172 11 ~ I. Homans, G.C. 1950. The Human Group. New York: Harcourt Brace. 12. Lott, A.~. & Lott, B.E. 1965. "Group cohesiveness as interpersonal attraction: A review of relationships with antecedent and consequent variables." Pay Bu! 64:259-309. 13. Fararo, T.~. & Skvoretz, J. 1987. "Unification research programs: Integrating two structural theories." Am Jour Soc 92: 1183- 1209. 14. Mayhew, Bruce H., J. Miller McPherson, Thomas Rotolo, and Lynn Sm~th-Lov~n. 1995. "Sex and Race Homogeneity In Naturally Occurring Groups." Social Forces 74:15-52. 15. Rosenbaum, M.E. 1986. "The repulsion hypothesis: On the non-development of relationships." Jour of Pers & Soc Psy 5 I: 1 1 56- 1 ~ 66. 16. Cartwright, D. & Zander, A. (1968) "Pressures to uniformity In groups: An introduction." Pp. 139-151 In Group Dynamics: Research anc! Theory, edited by Down Cartwright and Alvin Zander. New York NY: Harper & Row. 17. Fest~nger, L. 1950. "Informal social communication." Pay Rev 57:271-282. Ad. Schachter, S. 1951. "Deviation, rejection, and communication." Jour Abnormal Soc Psy 46:190-207. 19. FrieUk~n, N.E. 1984. "Structural cohesion and equivalence explanations of social homogeneity." Soc Meth & Res 12:235-261. 20. Marsden, P.V. & FrieUk~n, N.E. 1993. "Network studies of social influence." Soc Meth & Res 22:127-151. 21. Abelson, R.P. 1964. Mathematical Models of the Distribution of Attitudes Under Controversy. Pp. 142-160 in: Frederiksen, N. & Gulliken, H. (eds.~: Contributions to mathematical psychology. Holt, Rinehart & Winston, New York. 22. Harary, F. 1959. "A criterion for unanimity in French's theory of social power." In Stuclies in social power, edited by D. Cartwright, ~68-82. Ann Arbor: Institute for Social Research. 23. French, J. R. P. Jr. 1956. "The formal theory of social power." Psy Rev 63: ISI-94. 24. Heider, F. 1958. The Psychology of Interpersonal Relations. New York: Wiley. 25. Hopfield, John J. ~ 982. "Neural networks arid physical systems with emergent collective computational abilities." Proc NatAcacI Sci 79:2554-~. 26. Hopfield, J. J. and D. W. Tank. 1985. ""Neural" Computation of Decisions in Optimization Problems." Bio Cybernetics 52:141-152. 27. Hebb, D. O. (l 949), The Organisation of Behavior: A Neuropsychological Approach. John Wiley and Sons, New York, NY. 28. Nowak, A., Lewenste~n, M, & Tarkowski, W. ~ 993. "Repeller neural networks." Phys Rev 48: 1491-~. 29. Nowak, A. & Vallacher, R. ~997 "Computational social psychology: Cellular automata and neural network models of interpersonal dynamics." In S. Read and ~ Miller, eds. Connectionist Mociels of Social Reasoning and Social Behavior. Mahwah, NJ: Larence Eribaum. 30. Kitts, I.A., M.W. Macy, & Flache, A. 2000. "Structural learning: Attraction and conformity in task-oriented groups." Comp Math Org Theo~y,5: 129-145. 3~. Harsanyi, J. ~ 973. Games with randomly disturbed payoffs. International Journal of Game Theory 2:~-23. 32. Cartwright, D. & Harary, F. 1966. "Structural balance: A generalization of Heider's theo~." Pay Rev 63:277-93. 33. Davis, J.A. ~ 967. "Clustering and structural balance in graphs." Human Re!20:~-~87. DYNAMIC SOCIAL NETWORK AJODE~WG ED ISIS

12 34. Wasserman, S. & Faust, K. 1994. Social Network Analysis: Methods and Applications. Cambridge: Cambridge University Press. 35. Cartwright, D. & Gleason, T.C. 1966. "The number of paths and cycles in a digraph." P*ychometrika. 3 I: ~ 79- ~ 99. 36. Tajfel, H.& Turner, J.C. 1986. "The social identity theory of intergroup relations." in Psychology of Intergroup Relations, edited by S. Worche! and W. G. Austin. Chicago, Hi: Nelson-Hill DYNAMIC SOCIAL NETWORK A1ODEL~G~D TRYSTS 173

Next: Local Rules and Global Properties: Modeling the Emergence of Network Structure »
Dynamic Social Network Modeling and Analysis: Workshop Summary and Papers Get This Book
×
Buy Paperback | $73.00
MyNAP members save 10% online.
Login or Register to save!
Download Free PDF

In the summer of 2002, the Office of Naval Research asked the Committee on Human Factors to hold a workshop on dynamic social network and analysis. The primary purpose of the workshop was to bring together scientists who represent a diversity of views and approaches to share their insights, commentary, and critiques on the developing body of social network analysis research and application. The secondary purpose was to provide sound models and applications for current problems of national importance, with a particular focus on national security. This workshop is one of several activities undertaken by the National Research Council that bears on the contributions of various scientific disciplines to understanding and defending against terrorism. The presentations were grouped in four sessions – Social Network Theory Perspectives, Dynamic Social Networks, Metrics and Models, and Networked Worlds – each of which concluded with a discussant-led roundtable discussion among the presenters and workshop attendees on the themes and issues raised in the session.

  1. ×

    Welcome to OpenBook!

    You're looking at OpenBook, NAP.edu's online reading room since 1999. Based on feedback from you, our users, we've made some improvements that make it easier than ever to read thousands of publications on our website.

    Do you want to take a quick tour of the OpenBook's features?

    No Thanks Take a Tour »
  2. ×

    Show this book's table of contents, where you can jump to any chapter by name.

    « Back Next »
  3. ×

    ...or use these buttons to go back to the previous chapter or skip to the next one.

    « Back Next »
  4. ×

    Jump up to the previous page or down to the next one. Also, you can type in a page number and press Enter to go directly to that page in the book.

    « Back Next »
  5. ×

    To search the entire text of this book, type in your search term here and press Enter.

    « Back Next »
  6. ×

    Share a link to this book page on your preferred social network or via email.

    « Back Next »
  7. ×

    View our suggested citation for this chapter.

    « Back Next »
  8. ×

    Ready to take your reading offline? Click here to buy this book in print or download it as a free PDF, if available.

    « Back Next »
Stay Connected!