Varieties of homophily and core-periphery in metapopulation theories of cultural evolution

review
social networks
homophily
Author

Matthew A. Turner

Published

August 12, 2024

Intro

What do innovation in an online puzzle game (Derex and Boyd 2016; Moser and Smaldino 2023; Milzman and Moser 2023); health and weight loss interventions (Centola 2010a, 2011; Zhang and Centola 2019); the diffusion of ready-to-hand climate change adaptations (Jones, Ready, and Pisor 2021; Matthew A. Turner et al. 2023); prediction of HIV propagation through mutli-ethnic, urban social networks (Morris 1991); political polarization (Flache and Macy 2011; Matthew A. Turner and Smaldino 2018); and pro-democracy protest movements like the Arab Spring (Barberá et al. 2015; Steinert-Threlkeld 2017) all have in common? In all these cases, social network structure enabled more effective development and dissemination of collective intelligence (Centola 2022), potentially enabling more effective cultural adaptation (Galesic et al. 2023), but also sometimes amplifying problematic collective maladaptations in the case of polarization.

These examples show that network structure matters. However, these examples operationalize and quantify social network structure in different, idiosyncratic ways. It is therefore difficult to know whether the same network effects are at work in different study contexts. In turn, it is difficult to synthesize our scientific understanding of how network structure affects cultural adaptation in general. In this post I aim to synthesize two common approaches to quantifying social network structure observed to affect cultural adaptation: homophily (Kossinets and Watts 2009) and the degree of core-periphery structure in a network (Borgatti and Everett 1999; Priebe et al. 2019), which I dub core-peripheriness. I will first give brief introductions to the many possible definitions and operationalizations that exist for these two measures. Then to synthesize the two, I will fit a stochastic block model to social networks constructed using an algorithm I developed based on group-level homophily in the special case of a metapopulation (i.e., a population composed of sub-populations) composed of two groups.

This approach helps advance the social science of collective adaptation by initiating the development of a common framework for construction or measurement of social networks that could be used across formal, computational, and empirical studies. In this way we stop “playing twenty questions with nature” (Almaatouq et al. 2022) that happens when we have “twenty” (or more, or at least several) different ways of defining social network structure. A common framework can instead provide a foundation for new studies to build on. New studies can then multiply our knowledge instead of fracture it, as will continue to occur without a shared framework.

Approaches to operationalize social network structure

There are two main theoretical approaches to quantifying social network structure that I want to cover here: the homophily of the network and the core-peripheriness of the network. Homophily measures how frequently individuals interact with members of their own group compared to members of out-groups. In cultural evolutionary studies, homophily is almost always a global property, but we have ongoing work that relaxes this assumption to analyze asymmetric homophily, where homophily is different for each group. Core-peripheriness separates a metapopulation into one core group and a periphery composed of one or moregroups. The core group is characterized by dense connections between members with high transitivity, i.e., social triangles where if one individual is connected to two others, it is highly likely the two connections are also connected. The periphery is characterized by relatively sparse connections between individuals in the periphery and the core. It seems these are the two most fundamental approaches that give rise to what might be called higher-order properties of networks, such as transitivity and average path length, i.e., the mean number of social connections one must traverse to send a message from person to person like in the children’s game “telephone”.

Both of these network structure definitions themselves have been quantified or operationalized various ways. Here are some notes on how different papers quantified homophily and core-peripheriness.

Homophily

  • See (Kossinets and Watts 2009; Centola 2010b, 2011) for some modern model and empirical foundations. Evtushenko and Kleinberg (2021) define first- and second-order homophily of a node: first-order homophily is “individual tendency to link to similar others” and second-order homophily is “aggregate first-order homophily of its neighbors”.

    Outline:

    1. Types of Homophily: global (give empirical examples if possible, e.g. here, (Kossinets and Watts 2009; Centola 2010b, 2011; Golub and Jackson 2012)), second-order ((Evtushenko and Kleinberg 2021)), higher-order([@]; how to measure homophily? There is a difference between different measures of homophily and different forms of homophily: Kossinets & Watts also define homophily arising more from preference and those arising more from circumstance (work/profession, neighborhood, etc.), but of course a certain profession or neighborhood may be correlated with other potential group markers such as race or ethnicity.
  • Kossinets and Watts (2009)

  • Centola (2011)

  • Matthew A. Turner and Smaldino (2018)

Core-periphery

Core-periphery: - Milzman and Moser (2023) - Priebe et al. (2019)

The sophistication of culturally-evolved behaviors and the speed and success of cultural diffusion depends upon group structure in metapopulations of interest, where a metapopulation is a population with distinct groups. Group structure is an aggregate property of metapopulations that specifies the probabilities of a member of one group interacting with members of its own group or another group. My agent-based modeling work so far has operationalized group structure in terms of homophily, which is a measure of the frequency of within- and between-group interactions. An alternative operationalization defines group structure via latent parameters of Bayesian priors in a generative statistical model. These models are complementary, so that possibly fitting latent parameters that determine clustering and core-periphery structure in a stochastic block model have an equivalent representation in terms of asymmetric homophily. This points to an eventual mathematical explanation that could calculate stochastic block parameters directly from a given set of asymmetric homophily values.

  1. Stochastic block models two-group intro. Using our counting method for calculating homophily in those models.
  2. Different numbers of connections between groups.
  3. Fit stochastic block model to generated networks using our code via Cody Ross’s

Asymmetric homophily-defined networks fit to stochastic block models

To begin to understand the connection between asymmetric homophily and core-peripheriness, I will first explain how to simulate minority-majority social networks using group-level asymmetric homophily, then demonstrate how to fit a stochastic block model network with two blocks to infer the core-peripheriness of networks with different homophily levels. I close the section with a brief analysis that compares core-peripheriness of networks produced with different asymmetric homophily values.

Asymmetric homophily-defined networks

Homophily sets the probability that an agent potentially receives information from another agent within its own group, or from another group. The value is the number of percentage points more likely it is that a focal agent connects with in-group members compared to out-group members, with negative homophily indicating out-group connections are more likely than in-group.

When there is a minority group, teachers in the minority group are over-represented as teachers, indicated by the dense pink edges pointing from the minority to the majority, representing minority-teacher, majority-learner relationships.

Stochastic block model networks and core-peripheriness

The stochastic block model (SBM) uses probabilities of what we conceptualize as teacher-learner relationships between core and periphery to define the network. In SBM language, each group is a “block”. The log-odds probability of an edge from a node in group \(i\) to a node in group \(j\) is written \(B_{ij}\). With two groups, or “blocks”, there are thus four probabilities to define. To continue with the minority-majority framing from the asymmetric homophily model, we write these terms SBM connection probabilities as follows. \(B_{\mathrm{maj,maj}}\) is the connection probability among the core majority group; \(B_{\mathrm{min,maj}}\) is the probability of a minority member possibly teaches a majority member; \(B_{\mathrm{maj,min}}\) is the probability a majority member possibly teaches a minority member; and \(B_{\mathrm{min,min}}\) is the probability a minority member possibly teaches another minority member. To simplify notation and match other notation choices in the literature, we can call \(a=B_\mathrm{maj,maj}\), \(b = B_\mathrm{maj, min}\), \(c = B_\mathrm{min,maj}\), and \(d=B_\mathrm{min,min}\). These are then organized into the block matrix, \[ B = \begin{pmatrix} a & b \\ c & d \end{pmatrix}. \]

From this we can define a measure of core-peripheriness appropriate for our purposes, based on some approaches from the literature. For now we adapt the approach of Milzman and Moser (2023) who developed an agent-based model to better understand why participants in partially-connected groups outperformed participants in fully-connected groups in a cumulative cultural puzzle game in Derex and Boyd (2016). Milzman and Moser set \(b = c\) and use \(c\) in place of \(d\), but we will adapt their classifications to relax this assumption and assume that in general \(b \neq c\).

Milzman and Moser group core-peripheriness in one of three classes with fuzzy boundaries based on the relative values of the block connection probabilities. They adapt similar definitions in Priebe et al. (2019), who used the definitions to understand commonalities and differences in the information obtained through different methods of network dimensionality reduction. First, decentralized core-periphery structure has a densely-connected core but weakly connected periphery, and weak connectivity between the periphery and the core, i.e., \(a >> b,c,d\). Second, centralized core-periphery structure has more connections between the periphery and the core, but weak connectivity within the periphery, i.e., \(a,c,b >> d\). Finally, affinity structure has little core-periphery structure, and is dominated by within-group connections so that the diagonal elements of the block matrix dominate, i.e., \(a,d >> b,c\). As in our asymmetric homophily model, Milzman and Moser fix the total number of edges in the network, but vary \(a,b,c,d\) to create different networks that fall along the decentralized-centralized-affinity spectrum. They find that decentralized core-periphery structure optimizes cumulative cultural evolution in a simulated version of the experimental task in Derex and Boyd (2016).

For directed graphs, Elliott et al. (2020) provide the most thorough core-peripheriness measure, which they call core–periphery modularity, which “measures partition quality” . We will not use Elliot and co-authors definition exactly because their goal is to identify core-periphery structure, and there exist alternative measures which we build on below. First, just note that Elliot and co-authors generalize the now-classic treatment of core-periphery structure by Borgatti and Everett (1999), which only considered undirected graphs. Maximal core-peripheriness, according to Borgatti and Everett, is the situation where the core is fully connected, and every member of the periphery is connected to every core member, but not connected to any other periphery members. A future post may test Elliott and co-author’s core-periphery modularity for measuring core-peripheriness.

SBM parameters of asymmetric homophily-defined networks

Since decentralized core-periphery structure outperformed other structures, we will make an ad-hoc measure of core-peripheriness based on the fact that \(a >> b,c,d\) implies decentralized core-periphery. Let’s define core-peripheriness to thus be

\[ \Gamma = \frac{a}{b + c + d} \]

References

Almaatouq, Abdullah, Thomas L. Griffiths, Jordan W. Suchow, Mark E. Whiting, James Evans, and Duncan J. Watts. 2022. Beyond Playing 20 Questions with Nature: Integrative Experiment Design in the Social and Behavioral Sciences.” Behavioral and Brain Sciences. https://doi.org/10.1017/S0140525X22002874.
Barberá, Pablo, Ning Wang, Richard Bonneau, John T. Jost, Jonathan Nagler, Joshua Tucker, and Sandra González-Bailón. 2015. The critical periphery in the growth of social protests.” PLoS ONE 10 (11): 1–15. https://doi.org/10.1371/journal.pone.0143611.
Borgatti, Stephen P., and Martin G. Everett. 1999. Model of core / periphery structures.” Social Networks 21: 375–95.
Centola, Damon. 2010a. The spread of behavior in an online social network experiment.” Science 329 (5996): 1194–97. https://doi.org/10.1126/science.1185231.
———. 2010b. The Spread of Behavior in an Online Social Network Experiment.” Science 329 (5996): 1194–97. https://doi.org/10.1126/science.1185231.
———. 2011. An experimental study of homophily in the adoption of health behavior.” Science 334 (6060): 1269–72. https://doi.org/10.1126/science.1207055.
———. 2022. The network science of collective intelligence.” Trends in Cognitive Sciences 26 (11): 923–41. https://doi.org/10.1016/j.tics.2022.08.009.
Centola, Damon, Juan Carlos González-Avella, Víctor M Eguíluz, and Maxi San. 2007. Homophily, Cultural Drift, and the Co-Evolution of Cultural Groups.” Journal of Conflict Resolution 51 (6): 905–29.
Centola, Damon, and Michael Macy. 2007. Complex Contagions and the Weakness of Long Ties.” American Journal of Sociology 113 (3): 702–34.
Derex, Maxime, and Robert Boyd. 2016. Partial connectivity increases cultural accumulation within groups.” Proceedings of the National Academy of Sciences of the United States of America 113 (11): 2982–87. https://doi.org/10.1073/pnas.1518798113.
Efferson, Charles, Sonja Vogt, and Ernst Fehr. 2020. The promise and the peril of using social influence to reverse harmful traditions.” Nature Human Behaviour 4 (1): 55–68. https://doi.org/10.1038/s41562-019-0768-2.
Elliott, Andrew, Angus Chiu, Marya Bazzi, Gesine Reinert, and Mihai Cucuringu. 2020. Core-periphery structure in directed networks: Core-Periphery in Directed Networks.” Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences 476 (2241). https://doi.org/10.1098/rspa.2019.0783.
Evtushenko, Anna, and Jon Kleinberg. 2021. The paradox of second-order homophily in networks.” Scientific Reports 11 (1): 1–10. https://doi.org/10.1038/s41598-021-92719-6.
Flache, Andreas, and Michael W. Macy. 2011. Small Worlds and Cultural Polarization.” The Journal of Mathematical Sociology 35 (1-3): 146–76.
Galesic, Mirta, Daniel Barkoczi, Andrew M. Berdhal, Dora Biro, Giuseppe Carbone, Ilaria Giannoccaro, Robert L. Goldstone, et al. 2023. Beyond collective intelligence : Collective adaptation.” Journal of the Royal Society Interface 20 (20220736). https://doi.org/10.1098/rsif.2022.0736.
Golub, Benjamin, and Matthew O. Jackson. 2012. How homophily affects the speed of learning and best-response dynamics.” The Quarterly Journal of Economics, 1287–1338. https://doi.org/10.1093/qje/qjs021.Advance.
Jones, James Holland, Elspeth Ready, and Anne C. Pisor. 2021. Want climate-change adaptation? Evolutionary theory can help.” American Journal of Human Biology 33 (4): 1–17. https://doi.org/10.1002/ajhb.23539.
Kossinets, Gueorgi, and Duncan J. Watts. 2009. Origins of homophily in an evolving social network.” American Journal of Sociology 115 (2): 405–50. https://doi.org/10.1086/599247.
Milzman, Jesse, and Cody Moser. 2023. Decentralized core-periphery structure in social networks accelerates cultural innovation in agent-based model.” http://arxiv.org/abs/2302.12121.
Morris, Martina. 1991. A log-linear modeling framework for selective mixing.” Mathematical Biosciences 107 (2): 349–77. https://doi.org/10.1016/0025-5564(91)90014-A.
Moser, Cody, and Paul E. Smaldino. 2023. Innovation-facilitating networks create inequality.” Proceedings of the Royal Society B: Biological Sciences 290 (2011). https://doi.org/10.1098/rspb.2023.2281.
Priebe, Carey E., Youngser Park, Joshua T. Vogelstein, John M. Conroy, Vince Lyzinski, Minh Tang, Avanti Athreya, Joshua Cape, and Eric Bridgeford. 2019. On a two-truths phenomenon in spectral graph clustering.” Proceedings of the National Academy of Sciences of the United States of America 116 (13): 5995–6000. https://doi.org/10.1073/pnas.1814462116.
Steinert-Threlkeld, Zachary C. 2017. Spontaneous collective action: Peripheral mobilization during the arab spring.” American Political Science Review 111 (2): 379–403. https://doi.org/10.1017/S0003055416000769.
Turner, Matthew A, Alyson L Singleton, Mallory J Harris, Ian Harryman, Cesar Augusto Lopez, Ronan Forde Arthur, Caroline Muraida, and James Holland Jones. 2023. Minority-group incubators and majority-group reservoirs support the diffusion of climate change adaptations.”
Turner, Matthew A., and Paul E. Smaldino. 2018. Paths to Polarization: How Extreme Views, Miscommunication, and Random Chance Drive Opinion Dynamics.” Complexity. https://doi.org/10.1155/2018/2740959.
Vogt, Sonja, Nadia Ahmed Mohmmed Zaid, Hilal El Fadil Ahmed, Ernst Fehr, and Charles Efferson. 2016. Changing cultural attitudes towards female genital cutting.” Nature 538 (7626): 506–9. https://doi.org/10.1038/nature20100.
Zhang, Jingwen, and Damon Centola. 2019. Social Networks and Health: New Developments in Diffusion, Online and Offline.” Annual Review of Sociology 45 (1): 91–109. https://doi.org/10.1146/annurev-soc-073117-041421.