Contribution
It names and quantifies a variant of the friendship paradox. With a binary attribute (“active” or “inactive”), many nodes can find that more than half their neighbours are active although few nodes are. The paper shows how large the share of such nodes gets in synthetic and real network topologies as the correlation between degree and attribute, the degree assortativity and the heaviness of the degree distribution vary. It gives a statistical model (Eq 5) that predicts the share from the degree distribution, the joint degree distribution and the conditional attribute distribution. What is new is the measure and its dependence on assortativity, not the underlying mechanism, which is the friendship paradox’s: high-degree nodes are over-represented in neighbourhoods.
Key insight
Each node’s local view can be accurate and still unrepresentative. When the active nodes are the hubs, they sit in many neighbourhoods at once, so “a state that is globally rare in a network may be dramatically over-represented in the local neighborhoods of many individuals” (Abstract). Fig 1 makes the point with 14 nodes: the same graph with three active nodes gives every inactive node an active majority, or none, depending on which three. Nothing in the effect requires anyone to be loud, to misreport or to misperceive what their neighbours are. The error, where there is one, is in generalizing from neighbours to the population.
Assumptions
- Undirected graphs and a binary attribute. Directed follower graphs (Digg, Twitter) are made undirected by keeping only mutual links.
- The attribute’s placement is set by the experimenters. On both the synthetic and the real networks, a fraction P(x = 1) of nodes is activated and an attribute-swapping procedure raises the degree–attribute correlation ρ_kx to chosen values. The real networks supply topology only; no real attribute is observed on any of them.
- The paradox is measured at a threshold of one half. A node is “in the paradox regime” if more than half its neighbours are active (Eq 5, ϕ = 1/2).
- The model treats a node’s neighbours as independent draws with probability P(x′ = 1|k) of being active (Eq 4, a binomial), which depends only on the node’s degree through the joint degree distribution e(k, k′). For Poisson-like degree distributions a Gaussian approximation of P(x, k) is used; otherwise the empirical P(x, k) is plugged in.
Key results
- Scale-free synthetic networks (N = 10,000, p(k) ~ k^−α with α = 2.1, 2.4, 3.1; 5% active). “For α = 2.1, 60%–80% of the nodes will observe that more than half of their neighbors are active, even though only 5% of the nodes are, in fact, active.” The effect grows with ρ_kx, as the network becomes more disassortative (r_kk falls) and as the tail gets heavier (α falls) (Fig 2).
- Erdős–Rényi networks (⟨k⟩ = 5.2 and 2.5; 5%, 10% and 20% active). The effect is “much reduced compared to scale-free networks” but present, more so with more active nodes (Fig 3).
- Six real topologies (HepTh co-authorship, Reactome protein interactions, Digg, Enron email, Twitter, political blogs; r_kk from 0.27 to −0.22, Table 1), with 5–30% of nodes activated and ρ_kx raised by swapping. In the disassortative political blogs, Twitter and Enron networks, “as many as 60%–70% of nodes have more than half of their neighbors in the active state, even though only 20% of the nodes are active”. Less in Digg, and less again in HepTh and Reactome (Fig 4).
- The model. P(x′ = 1|k) = Σ_k′ P(x′ = 1|k′) e(k, k′)/q(k) (Eq 3), and the paradox share is P_{>1/2} = Σ_k p(k) Σ_{n>k/2} C(k, n) P(x′ = 1|k)^n [1 − P(x′ = 1|k)]^{k−n} (Eq 5). With the empirical P(x, k) it “does a good job explaining most observations”. The Gaussian approximation fits for α = 3.1 and deviates for α = 2.1 and for strongly assortative or disassortative networks (Fig 5).
- Assortativity alone is not enough. Two networks with the same p(k) and r_kk but different e(k, k′) “can display different amounts of the paradox” (S1 Fig, as reported in the text; S1 Fig not seen).
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | When an attribute is positively correlated with degree, the share of nodes with an active majority among their neighbours can greatly exceed the attribute’s global share | strong for the networks studied | simulation on synthetic networks and on six real topologies with assigned attributes (Figs 2–4) |
| C2 | The effect is stronger in disassortative and in heavier-tailed networks, and weaker but present in Poisson networks | strong for the networks studied | Figs 2–4; the sign argument from Eqs 1 and 3 |
| C3 | A binomial model built from p(k), e(k, k′) and P(x, k) predicts the share | moderate: fits the simulations, with an approximation that fails for heavy tails and strong (dis)assortativity | Eq 5; Figs 2–5 |
| C4 | The illusion can lead people to accept a globally rare behaviour as a norm and can help a contagion spread from few adopters | weak: an interpretation | the Fig 1 illustration with a threshold model; no contagion simulated and no person studied |
| C5 | It may explain why students overestimate peers’ drinking, and “apparently sudden shifts in public attitudes” (the Arab Spring, gay marriage) | weak: speculation | Introduction and Conclusion; no data on either |
Concepts
- Majority illusion: the paper’s term for nodes observing that more than half their neighbours are active when the active state is globally rare; its “magnitude” is “the fraction of nodes more than half of whose neighbors are active”.
- Friendship paradox and generalized friendship paradox: on average one’s neighbours have more neighbours (Feld 1991), and more of any attribute correlated with degree.
- Degree–attribute correlation ρ_kx: the Pearson correlation between a node’s degree and its binary attribute (Eq 2).
- Degree assortativity r_kk: the Pearson correlation between the degrees of connected nodes (Eq 1, Newman 2002).
- Class size bias: the paper’s general frame for the effect; popular classes and events are over-represented in what people experience.
Connections
It builds on the friendship paradox (Feld 1991) and its generalization to attributes (Eom & Jo 2014; Hodas, Kooti & Lerman 2013), on threshold models of contagion (Granovetter 1978; Watts 2002; Centola), and on Newman’s assortativity and rewiring procedure. It offers itself as “an alternate network-based mechanism” beside the psychologists’ accounts of social perception bias: false consensus (Marks & Miller 1987), selective exposure and selective disclosure (Kitts 2003), and pluralistic ignorance, citing Prentice & Miller 1993 (LIT-1370) as ref 31 and Miller & Prentice 1994. It also cites Centola, Willer & Macy’s “Emperor’s Dilemma” on self-enforcing unpopular norms, without engaging it.
Bearing on the record
- CLAIM-415 (a person organized by a judgement attributed to a
collective that does not hold it). The claim grounds itself on this
paper as one of the “mechanisms” for a perceived plurality with no real
one behind it. The reading supports that use only in part.
- Supports: a person’s local sample can show a majority where the population holds a small minority, with no misperception of anyone. That is a structural route to an attributed collective that does not exist as attributed, which is what the claim’s first step needs.
- Does not support: the claim’s own examples. The paper has no “loud minority” (the owner’s U39): its mechanism is degree, not volume or frequency of expression, and its active nodes are as visible as any other. It measures no belief, conformity or self-conception, and nothing about rejection or resistance (the claim’s second step). Whether anyone acts on the local majority is assumed in a toy threshold example (Fig 1), not shown.
- A distinction the claim’s A39 blurs. The paper treats pluralistic ignorance as one more “social perception bias” it may help explain. But the two are different errors. In pluralistic ignorance (in the record as List’s falsely attributed common attitude, NOTE-554) members misjudge what others hold. In the majority illusion each node sees its neighbours’ states correctly and the sample is skewed. The paper’s speculation about student drinking runs the two together. The claim should cite them as two mechanisms, not one.
- THEORY-242 (spiral of silence; the network’s direct distortion small). No contradiction, but the two should not be cited as if they measured the same thing. Vilone and Polizzi (NOTE-810) set themselves against this paper and find the topological term small. Their measure is the gap between the average local climate and the global expressed opinion, with the minority on nodes of one chosen degree. This paper’s measure is the share of nodes with an active local majority, with the attribute pushed onto hubs until ρ_kx is high, and it finds the share large only when ρ_kx is high and the network disassortative. Both can be true. THEORY-242’s “the network’s direct distortion of what members see is small” holds for its model, not for networks in general.
- CLAIM-355 (the constructed owner). Not touched. The paper says nothing about self-models.
- THEORY warranted? Yes, one: the dependence of the local-majority share on degree–attribute correlation, assortativity and tail weight is a substantive claim about networks the record does not hold. Filed as THEORY-887, scoped to perception of neighbours’ states and not to belief, conformity or contagion.
- No instruction for machine-learning practice; nothing for the anthology.
Limitations
- No real attribute is observed. On every network the authors choose which nodes are active and raise ρ_kx by swapping until it reaches the value they want. The paper shows that real topologies permit a large illusion; whether real attributes (opinions, drinking, adoption) are in fact correlated with degree strongly enough is not examined.
- No people. Nothing about what anyone perceives, believes or does. “Leading them to accept as a norm a behavior that is globally rare” is the authors’ interpretation (Conclusion).
- No contagion is simulated. The claim that the illusion “may accelerate the spread of social contagions” rests on the 14-node Fig 1.
- The threshold is fixed at one half, and the measure counts nodes, not how far each node’s local share departs from the global one.
- The Arab Spring and gay-marriage remarks are offered without data.
Open questions
- In a network with a real, measured attribute, how large is ρ_kx, and how many nodes are in the paradox regime? A study pairing observed attributes with observed topology would settle whether the illusion is common or only possible.
- Do people’s estimates of prevalence track their neighbourhood share, or do they discount for their neighbours’ popularity? A survey of perceived prevalence set beside ego networks would test the bridge from sample to belief the paper assumes.
- How does the illusion combine with selective disclosure and silence (THEORY-242)? A model with both a degree–attribute correlation and an option of silence would show whether they add or cancel.