Modeling limited attention in opinion dynamics by topological interactions

This work explores models of opinion dynamics with opinion-dependent connectivity. Our starting point is that individuals have limited capabilities to engage in interactions with their peers. Motivated by this observation, we propose a continuous-time opinion dynamics model such that interactions take place with a limited number of peers: we refer to these interactions as topological, as opposed to metric interactions that are postulated in classical bounded-confidence models. We observe that topological interactions produce equilibria that are very robust to perturbations.

READ FULL TEXT VIEW PDF
POST COMMENT

Comments

There are no comments yet.

Authors

page 1

page 2

page 3

page 4

03/20/2018

Asynchronous opinion dynamics on the k-nearest-neighbors graph

This paper is about a new model of opinion dynamics with opinion-depende...
02/21/2020

Hysteresis and disorder-induced order in continuous kinetic-like opinion dynamics in complex networks

In this work we tackle a kinetic-like model of opinions dynamics in a ne...
03/01/2019

Egocentric Bias and Doubt in Cognitive Agents

Modeling social interactions based on individual behavior has always bee...
04/30/2020

Extremism definitions in opinion dynamics models

There are several opinion dynamics models where extremism is defined as ...
02/06/2021

Opinion Dynamics Incorporating Higher-Order Interactions

The issue of opinion sharing and formation has received considerable att...
10/08/2018

Noise-synchronizability of opinion dynamics

With the analysis of noise-induced synchronization of opinion dynamics w...
12/03/2021

Investigating Opinion Dynamics Models in Agent-Based Simulation of Energy Eco-Feedback Programs

According to research, reducing consumer energy demand through behaviour...
This week in AI

Get the week's most popular data science and artificial intelligence research sent straight to your inbox every Saturday.

1 Introduction

Driven by the evolution of digital communication and social networking services, there is an increasing interest for mathematical models of opinion dynamics in social networks. Among the many models proposed in the literature, a few have become popular in the control community [27, 26]. In the perspective of the control community, opinion dynamics distinguish themselves from consensus dynamics because consensus is prevented by some other feature of the dynamics. In many popular models, this feature is an opinion dependent limitation of the connectivity. Chief examples are bounded confidence models [22, 16], where social agents influence each other iff their opinions are closer than a threshold.

This way of defining influence assumes that agents have always access to the opinions of all fellow agents and may lead to agents being influenced by a large number of their fellows, possibly the whole population. Instead, the number of possible interactions is bounded in practice by the limited time and efforts that individuals can devote to social interactions. Similar limitations of attention are well documented in psychology and sociology, for instance by the notion of Dunbar number [18], and become evermore crucial in today’s age of information bonanza [21]. Indeed, in online social media, natural limitations of attention interplay with the way the online platforms are designed. Users interact via the contents they share: out of the pool of fresh contents, the online platform selects for each user the best contents in order to maximize engagement, mainly based on similarities between users [23]. As the notion of Dunbar number was originally defined with reference to primates, the reader will not find surprising that similar ideas have also been fundamental in the study of flocking in animal groups, as testified by numerous theoretical and experimental works [5, 20, 15, 24, 3]. The importance of considering networks where the number of neighbors is limited has also been understood by graph theorists, who have studied the properties of what they call -nearest-neighbors graph: for instance, it is known that must be logarithmic in to ensure connectivity [4].

However, few works have incorporated this important observation in suitable models of opinion dynamics. Before surveying these important references, we briefly describe the contribution of this paper. In our effort to make the case for limited attention in opinion dynamics, we study a simple continuous-time model (first appeared in the survey paper [2]) in which every agent is influenced by her closest nearest neighbors. In this paper, we provide some preliminary results about this continuous-time dynamics. Our results concentrate on two axes: studying the main properties of its equilibria, including their robustness to disruptions, and proving convergence results in special cases. We describe the equilibria of the dynamics, distinguishing a special type of clusterization equilibria that are constituted of separate clusters, and we discuss the robustness of clustered equilibria to disruptions, such as the addition of new agents. Regarding the question of convergence, we are able to provide a proof in two cases: when the total number of agents is small enough compared to number of neighbors , namely ; and when , that is, agents are only influenced by one “best friend”.

The difficulties in studying -nearest-neighbors dynamics originate from two key features: (1) interactions are not reciprocal; (2) whether two agents interact does not only depend on their two states, but also on the states of all the other agents. In the literature, models with any of these features are still relatively few. In classical bounded confidence models [22, 16], interactions are reciprocal as long as the interaction thresholds are equal for all agents [22, 7, 8, 10, 9], and any lack of reciprocity makes the analysis much more delicate [25, 12, 14]. In our model, not only interactions are non-reciprocal, but they are also non-metric: whether two agents interact is not solely determined by the distance between their two opinions. For this reason, we follow a consolidated tradition [5] and refer to our connectivity model where agents can interact with their nearest neighbors as topological.

Topological interactions are becoming increasingly popular in the applied mathematics community, especially for second order models [13]. Kinetic and continuum models with topological interactions are also actively studied [6, 31, 17]. Among first order “opinion” models, [1] has recently used Petri nets to define a class of models where interactions depend on the opinions of multiple agents: despite some similarities, our model does not belong to this class. In our recent papers [29, 28], we have studied two dynamics with asynchronous updates (with and without sub-sampling, respectively) that are discrete-time counterparts of the model we propose here. Finally, our contribution here differs from the one of [2] as the latter focuses on specific properties of the equilibria, such as the distribution of their clusters’ sizes, studied by extensive simulations, whereas we are interested in analytical results about dynamical properties like convergence to the equilibria and about their robustness to perturbations. Our robustness analysis is inspired by the approach taken in [7] for bounded confidence models.

The rest of this paper has the following structure. Section 2 introduces the model, Section 3 develops its analysis, and Section 4 discusses our results.

2 Mathematical model

Let and be two integers with

and let be the set of agents. Each agent is endowed with a scalar opinion . For every agent , her neighborhood is defined in the following way. The elements of are ordered by increasing values of ; then, the first elements of the list (i.e. those with smallest distance from ) form the set of current neighbors of . Should a tie between two or more agents arise, priority is given to agents with lower index. Based on the current definition of , agent ’s opinion evolves according to

(1)

In order to describe the inter-agent interactions allowed by a state , it is convenient to define the directed graph

where is the set of neighbors of . Clearly, if the graph is complete (up to self-loops). In using some simple graph theory in this paper, we take some background and standard jargon for granted: a concise summary can be found in [19, Ch. 1].

The chosen tie-breaking rule makes the right-hand side well defined for any . The neighborhoods depend on the current state and, therefore, on time. This fact makes dynamics (1) a piecewise-continuous system [11]. Its solutions shall be intended in a semi-classical sense, that is, as piecewise-smooth solutions such that the right-derivative of the solution is equal to the right-hand side at all times, that is,

We conjecture that a forward complete and unique solution exists from every initial condition: a rigorous verification of this fact, which is assumed to hold true in the rest of this paper, is left to future work. Note that choosing a more general notion of solutions, e.g. Caratheodory’s, would prevent unicity and the produced multiple solutions would make the subsequent analysis more delicate.

3 Analysis

This section details our results on dynamics (1). We first study equilibria, then convergence properties, and finally reconsider equilibria to study their stability.

3.1 Equilibria

A cluster is a subset of agents that have the same opinion: such that for all . A state is called clusterization if every agent belongs to a cluster of at least elements. Finally, a clusterization with only one cluster is said to be a consensus.

A state is said to be an equilibrium for (1) when the right-hand side is zero. For any , it is immediate to see that if belongs to a cluster of at least elements. This condition is also necessary for and . Therefore, all clusterizations are equilibria and all non-consensus equilibria have at least two clusters of at least elements, but not all equilibria are clusterizations. It is possible to obtain a simple counterexample by considering and with

(2)

Note that this example exploits the tie-breaking rule. However, this is not necessary, as the following example shows: consider and with

3.2 Dynamical properties

We can readily observe that, for any two agents and ,

(3)

This formula allows us to derive a few consequences. First, we observe that if for all , then . Second, we can deduce that the dynamics preserves the order of the agents.

Proposition 1 (Order preservation)

If then for all .

Proof

Observe that (3) can be rewritten as

where the inequality holds because and for any and . By this bound and Gronwall lemma, cannot reach zero in finite time. ∎

As a consequence of this property, we can assume from now on with no loss of generality that the agents are sorted in ascending order of opinions, that is, for all and all . The following proposition formally justifies this fact.

Proposition 2 (Re-ordering agents)

Let be a solution and be a permutation on the index set . Assume that for all pairs of distinct indices the permutation satisfies if either or and . Then, the following facts hold true:

  1. for all and for all ;

  2. if , then for all .

Proof

To verify the first claim, notice that the definition of does not interfere with the tie-breaking rule that is used in the definition of the neighborhoods, therefore and the dynamics of the agent that before the permutation had index is unchanged.

To verify the second claim, observe the following facts. If , then for by Proposition 1. If , then and therefore also for by (3).∎

Finally, we can deduce a convergence result for small groups.

Proposition 3 (Consensus for small groups)

If , then converges to a consensus.

Proof

Since , the two agents with lowest and highest opinion share at least one neighbor. Therefore, their difference evolves according to

which implies exponential convergence to zero by Gronwall lemma. ∎

3.3 Special case

In the case , the dynamics takes the form

where denotes the index of the closest agent to . This specific form has three important consequences.

Lemma 1

If , then the following facts hold true.

  1. All equilibria are clusterizations.

  2. For every , the interaction graph is the union of weakly connected components, such that each component contains exactly one circuit of length 2 and the two nodes of the circuit can be reached from all nodes of the component.

  3. Two disconnected components cannot become connected in the evolution.

Proof

Claim 1: We observe that the only possibility for the right-hand side to be zero is that for all .

Claim 2: Observe that can only be equal to either or , except for the extreme agents, for which necessarily and . Therefore, the sequence is such that and

and must therefore change sign an odd number of times. Where it changes from positive to negative, there is a pair of reciprocal edges; where it changes from negative to positive, there is a disconnection. Therefore, every connected component has a pair of nodes that are connected to each other and that can be reached through a directed path from all other nodes. See Figure 

1 for an illustrative example.

Claim 3: Let there be a disconnection between and . Then,

implying that the distance cannot decrease. Moreover,

because the second term either is negative or, if positive, must be smaller or equal in magnitude than . Therefore, cannot increase. Since an analogous reasoning implies that cannot increase, the two components cannot become connected in the future. ∎

Figure 1: Example of weakly connected component of graph .

These facts allow to draw a conclusion about convergence.

Proposition 4 (Clusterization)

If , then converges to a clusterization.

Proof

The third statement of Lemma 1 implies that weakly connected component can only split. Since the number of individuals is finite, the splitting process terminates with a finite number of constant weakly connected components. After that termination time, the topology does not change. Since each connected component has a globally reachable node, then each group of agents is guaranteed to converge to consensus [19, p. 61], therefore producing a clusterization. ∎

Unfortunately, the idea of the proof of Proposition 4 does not extend to , because in that case disconnected components can become connected.

3.4 Stability and robustness of equilibria

It is easy to see that non-clusterization equilibria are not stable in general. For instance, consider example (2) with a small perturbation on agent 7: the ensuing dynamics leads to a clusterization with two clusters. Instead, clusterizations exhibit several stability properties.

We shall begin by considering small perturbations of the opinions. We say that a clusterization state is structurally stable if, after a perturbation, the dynamics converges to another clusterization that has the same clusters (though not necessarily taking on the same opinion values). More formally, the clusterization is said to be structurally stable if there exists a neighborhood of such that, for every in that neighborhood, any solution issuing from converges to a clusterization that has the following property: for any pair of individuals, if and only if .

Proposition 5 (Structural stability of small clusters)

A clusterization is structurally stable if and only if all of its clusters have cardinality not larger than .

Proof

If all clusters have cardinality not larger than , then after the perturbation Proposition 3 can be applied. To prove the opposite implication, observe that if one cluster has cardinality at least , then a suitable perturbation can split it into two separate clusters of cardinality at least , thereby creating a clusterization with different structure.∎

We shall also consider different kinds of disruptions, namely the addition or removal of one agent. We say that a clusterization is stable to these disruptions if, after the addition or removal of an agent, the other agents do not change their opinion.

Proposition 6 (Stability to removals)

A clusterization is stable to removals if and only if all of its clusters have cardinality larger than .

Proof

It is clear that agents in a cluster remain at equilibrium after the removal, unless the cluster size goes below the threshold . ∎

Proposition 7 (Stability to additions)

Every clusterization is stable to additions.

Proof

Agents within a cluster of size at least will not be influenced by any new arrival. ∎

4 Conclusion

The stability properties of the equilibria of dynamics (1) should be contrasted with the lack thereof shown by the equilibria of the corresponding metric bounded confidence model, which reads as

(4)

where is an interaction radius. It is well-known [8, 10] that this dynamics converges to clusterizations. If a new agent is added to such a clusterization, either the new agent is too far apart from the original agents and nothing happens, or the new agent falls within the visibility radius from a cluster. In the latter case, the new agents and the agents in the cluster influence each other and therefore change their opinions, converging to a common intermediate value. Actually, if the new agent falls within the visibility radius of two clusters, the two clusters eventually merge.

In contrast, clusters produced by (1) are much more stable. In our opinion, this stability intriguingly reminds the stability that is exhibited by norms and organizations in societies. Indeed, sociologists and ethologists have observed since a long time [32, 33, 30] that social norms and social structures are typically rather stable across time, despite the fact that the composition of the social groups evolve, notably with the arrival of new members. Our insights about -neighbor interactions suggest that limitations of attention can have stabilizing effects in societies.

Acknowledgements

The authors are grateful to Emiliano Cristiani, Julien Hendrickx, Samuel Martin, Benedetto Piccoli and Tommaso Venturini for fruitful discussions and collaborations that, along the years, have shaped their point of view on the topics of this paper.

References

  • [1] Angeli, D., Manfredi, S.: A Petri net approach to consensus in networks with joint-agent interactions. Automatica 110, 108466 (2019)
  • [2] Aydoğdu, A., Caponigro, M., McQuade, S., Piccoli, B., Pouradier Duteil, N., Rossi, F., Trélat, E.: Interaction network, state space, and control in social dynamics. In: Bellomo, N., Degond, P., Tadmor, E. (eds.) Active Particles, Volume 1: Advances in Theory, Models, and Applications, pp. 99–140. Springer (2017)
  • [3] Aydoğdu, A., Frasca, P., D’Apice, C., Manzo, R., Thornton, J., Gachomo, B., Wilson, T., Cheung, B., Tariq, U., Saidel, W., Piccoli, B.: Modeling birds on wires. Journal of Theoretical Biology 415, 102–112 (2017)
  • [4]

    Balister, P., Bollobás, B., Sarkar, A., Walters, M.: Connectivity of random k-nearest-neighbour graphs. Advances in Applied Probability

    37(1), 1–24 (2005)
  • [5] Ballerini, M., Cabibbo, N., Candelier, R., Cavagna, A., Cisbani, E., Giardina, I., Lecomte, V., Orlandi, A., Parisi, G., Procaccini, A., Viale, M., Zdravkovic, V.: Interaction ruling animal collective behavior depends on topological rather than metric distance: Evidence from a field study. Proceedings of the National Academy of Sciences 105(4), 1232–1237 (2008)
  • [6] Blanchet, A., Degond, P.: Topological interactions in a Boltzmann-type framework. Journal of Statistical Physics 163(1), 41–60 (Apr 2016)
  • [7] Blondel, V.D., Hendrickx, J.M., Tsitsiklis, J.N.: On Krause’s multi-agent consensus model with state-dependent connectivity. IEEE Transactions on Automatic Control 54(11), 2586–2597 (2009)
  • [8] Blondel, V.D., Hendrickx, J.M., Tsitsiklis, J.N.: Continuous-time average-preserving opinion dynamics with opinion-dependent communications. SIAM Journal on Control and Optimization 48(8), 5214–5240 (2010)
  • [9] Canuto, C., Fagnani, F., Tilli, P.: An Eulerian approach to the analysis of Krause’s consensus models. SIAM Journal on Control and Optimization 50(1), 243–265 (2012)
  • [10] Ceragioli, F., Frasca, P.: Continuous and discontinuous opinion dynamics with bounded confidence. NonLinear Analysis and its Applications B 13(3), 1239–1251 (2012)
  • [11] Ceragioli, F., Frasca, P.: Discontinuities, generalized solutions, and (dis)agreement in opinion dynamics. In: Tarbouriech, S., Girard, A., Hetel, L. (eds.) Control Subject to Computational and Communication Constraints: Current Challenges, pp. 287–309. Springer International Publishing (2018)
  • [12] Chazelle, B., Wang, C.: Inertial Hegselmann-Krause systems. IEEE Transactions on Automatic Control 62(8), 3905–3913 (Aug 2017)
  • [13] Chen, C., Chen, G., Guo, L.: On the minimum number of neighbors needed for consensus of flocks. Control Theory and Technology 15(4), 327–339 (2017)
  • [14] Chen, G., Su, W., Mei, W., Bullo, F.: Convergence properties of the heterogeneous Deffuant-Weisbuch model. arXiv preprint arXiv:1901.02092 (2019)
  • [15] Cristiani, E., Frasca, P., Piccoli, B.: Effects of anisotropic interactions on the structure of animal groups. Journal of Mathematical Biology 62(4), 569–588 (2011)
  • [16] Deffuant, G., Neau, D., Amblard, F., Weisbuch, G.: Mixing beliefs among interacting agents. Advances in Complex Systems 03(01n04), 87–98 (2000)
  • [17] Degond, P., Pulvirenti, M.: Propagation of chaos for topological interactions. The Annals of Applied Probability 29(4), 2594–2612 (2019)
  • [18] Dunbar, R.: Neocortex size as a constraint on group size in primates. Journal of Human Evolution 22(6), 469–493 (1992)
  • [19] Fagnani, F., Frasca, P.: Introduction to Averaging Dynamics over Networks. Lecture Notes in Control and Information Sciences, Springer Nature (2017)
  • [20] Giardina, I.: Collective behavior in animal groups: Theoretical models and empirical studies. HFSP Journal 2(4), 205–219 (2008), pMID: 19404431
  • [21] Gonçalves, B., Perra, N., Vespignani, A.: Modeling users’ activity on Twitter networks: Validation of Dunbar’s number. PloS one 6(8) (2011)
  • [22] Krause, U.: A discrete nonlinear and non-autonomous model of consensus formation. Communications in Difference Equations pp. 227–236 (2000)
  • [23] Lazer, D.: The rise of the social algorithm. Science 348(6239), 1090–1091 (2015)
  • [24] Martin, S.: Multi-agent flocking under topological interactions. Systems & Control Letters 69, 53–61 (2014)
  • [25] Mirtabatabaei, A., Bullo, F.: Opinion dynamics in heterogeneous networks: Convergence conjectures and theorems. SIAM Journal on Control and Optimization 50(5), 2763–2785 (2012)
  • [26] Proskurnikov, A., Tempo, R.: A tutorial on modeling and analysis of dynamic social networks. Part II. Annual Reviews in Control 45, 166–190 (2018)
  • [27] Proskurnikov, A.V., Tempo, R.: A tutorial on modeling and analysis of dynamic social networks. Part I. Annual Reviews in Control 43, 65–79 (Mar 2017)
  • [28] Rossi, W.S., Frasca, P.: Asynchronous opinion dynamics on the -nearest-neighbors graph. In: IEEE Conference on Decision and Control. pp. 3648–3653 (Dec 2018)
  • [29] Rossi, W.S., Frasca, P.: Opinion dynamics with topological gossiping: Asynchronous updates under limited attention. IEEE Control Systems Letters 4(3), 566–571 (2020)
  • [30] Sethi, R.: Evolutionary stability and social norms. Journal of economic behavior & organization 29(1), 113–140 (1996)
  • [31] Shvydkoy, R., Tadmor, E.: Topological models for emergent dynamics with short-range interactions. arXiv preprint (2018)
  • [32] Simmel, G.: The persistence of social groups. American Journal of Sociology 3(5), 662–698 (1898)
  • [33] Van de Waal, E., Borgeaud, C., Whiten, A.: Potent social learning and conformity shape a wild primate’s foraging decisions. Science 340(6131), 483–485 (2013)