Source-linked AI summary
The Majority Illusion in Social Networks
Kristina Lerman, Xiaoran Yan, Xin-Zeng Wu
TL;DR
The paper asks how social-network structure can make globally rare behaviors appear locally common, affecting perceptions and contagion. It analyzes the majority illusion in synthetic and real-world networks and develops a statistical model for its magnitude. The effect is stronger when highly connected nodes are active and in networks with heterogeneous degree distributions and suitable degree correlations.
Problem
Social behaviors spread through local imitation, but people estimate global prevalence from friends whose behaviors may be systematically overrepresented in their neighborhoods.
Method
The paper studies binary active attributes in synthetic and real-world networks and develops a statistical model incorporating network structure and attribute configuration.
Results
The majority illusion is stronger when better-connected nodes are active, when degree distributions are heterogeneous, and under network structures with amplifying degree correlations.
Takeaways & Limitations
A globally rare behavior may appear locally prevalent, potentially facilitating social contagions and contributing to systematic misperceptions of risky behavior.
Abstract
from arXiv · showhide
Social behaviors are often contagious, spreading through a population as individuals imitate the decisions and choices of others. A variety of global phenomena, from innovation adoption to the emergence of social norms and political movements, arise as a result of people following a simple local rule, such as copy what others are doing. However, individuals often lack global knowledge of the behaviors of others and must estimate them from the observations of their friends' behaviors. In some cases, the structure of the underlying social network can dramatically skew an individual's local observations, making a behavior appear far more common locally than it is globally. We trace the origins of this phenomenon, which we call "the majority illusion," to the friendship paradox in social networks. As a result of this paradox, a behavior that is globally rare may be systematically overrepresented in the local neighborhoods of many people, i.e., among their friends. Thus, the "majority illusion" may facilitate the spread of social contagions in networks and also explain why systematic biases in social perceptions, for example, of risky behavior, arise. Using synthetic and real-world networks, we explore how the "majority illusion" depends on network structure and develop a statistical model to calculate its magnitude in a network.
I. INTRODUCTION
The paper introduces the majority illusion: network structure can make globally rare behaviors appear common in many individuals’ local neighborhoods. It links this effect to the friendship paradox and examines its implications for social contagions.
- The friendship paradox means that people’s friends tend to have more friends than they do, creating systematically biased local observations.This degree imbalance has also informed vaccination, social intervention, and outbreak-detection strategies.
- The paper extends this paradox to binary attributes, labeling nodes with the attribute active and the others inactive.Examples include purchasing an iPhone or having red hair.
- Under some conditions, many nodes observe most of their neighbors as active even when the attribute is globally rare, producing the majority illusion.The effect arises because highly connected nodes are overrepresented in local neighborhoods.
- With an activation threshold of φ = 0.5, identical network topologies can produce complete activation in one configuration but no activation in another.The difference is caused by how active nodes are positioned, which changes perceived neighborhood prevalence.
- The paper measures paradox strength as the fraction of nodes with a majority of active neighbors and studies its dependence on network structure and active-node configuration.It uses synthetic and real-world networks, empirical analysis, and theoretical analysis.
II. RESULTS
The paper characterizes how degree structure, degree correlations, and degree–attribute correlations shape the majority illusion. It formalizes network and attribute configurations using degree and joint distributions, while preserving degree distributions during rewiring and tuning attribute correlations through swaps.
- Network structure: The neighbor degree distribution is q(k) = kp(k)/⟨k⟩, because high-degree nodes are over-represented among randomly reached neighbors.Here, p(k) describes node degrees and ⟨k⟩ is the average node degree.
- Network structure: Degree correlations are represented by e(k, k′), while assortativity r measures the Pearson correlation between degrees at opposite ends of edges.Positive r indicates links between similar degrees; negative r indicates links between dissimilar degrees.
- Network structure: Edge rewiring changes degree assortativity without changing the degree distribution p(k).The procedure swaps edges between selected pairs until a desired assortativity is achieved.
- Attribute configuration: Binary attributes are encoded by P(x, k), with x = 1 denoting active nodes and x = 0 denoting inactive nodes; ρkx measures degree–attribute correlation.The average degree of active nodes, ⟨k⟩x=1, and the standard deviations σk and σx enter the correlation calculation.
- Attribute configuration: Attribute swapping adjusts ρkx by exchanging active and inactive labels according to their node degrees.To increase ρkx, the procedure swaps an active lower-degree node with an inactive higher-degree node.
A. “Majority Illusion” in Synthetic and Real-world Networks
Synthetic and real-world networks show that the majority illusion can affect many nodes even when active nodes are globally rare. Its magnitude increases with degree–attribute correlation, disassortativity, and heavier-tailed degree distributions, and it also appears in observed networks.
- Synthetic networks: The majority illusion strengthens as degree–attribute correlation increases, assortativity decreases, and the degree distribution becomes heavier-tailed.A smaller degree-distribution exponent α corresponds to a heavier-tailed distribution.
- Synthetic networks: Even with α = 3.1, some conditions produce a substantial fraction of nodes experiencing the majority illusion.Theoretical lines in Figure 2 estimate the paradox using Equation 4.
- Synthetic networks: The majority illusion also occurs in Erdős–Rényi networks with Poisson degree distributions, although it is much reduced compared with scale-free networks.These experiments used networks of 10,000 nodes and activated 5%, 10%, or 20% of nodes.
- Real-world networks: In the political blogs network, 60%–70% of nodes may have a majority of active neighbors when only 20% of nodes are active and degree–attribute correlation is sufficiently high.The effect also appears in Digg and, to a lesser degree, HepTh; positive assortativity reduces its magnitude.
B. Modeling “Majority Illusion” in Networks
The paper models majority-illusion strength as the fraction of nodes whose neighbors exceed an activation threshold, using degree structure, degree correlations, and attribute configuration. The model captures empirical observations while showing that detailed joint degree structure can matter beyond global assortativity.
- Empirical evaluation: In Erdős-Rényi-type networks, Figure 3 varies magnitude with ρkx, assortativity r, mean degree, and the fraction of active nodes.The networks contain 10,000 nodes, with mean degree 5.2 in the top row and 2.5 in the bottom row.
- Empirical evaluation: In real-world networks, Figure 4 varies majority-illusion magnitude with degree–attribute correlation ρkx and active-node fraction P(x = 1).The plotted lines are calculations from the statistical model.
- Network structure: The probability of observing an active neighbor depends on degree correlations through e(k, k′), with stronger paradoxes expected in disassortative networks.The relevant expressions weight the joint degree distribution differently from the assortativity calculation.
- Statistical model: The model calculates the probability that a degree-k node has more than fraction φ of its neighbors active, then aggregates this probability across nodes.The paper focuses on φ = 1/2, representing a straight majority.
- Statistical model: Equation 4 applies when the degree sequence, joint degree distribution e(k, k′), and conditional attribute distribution P(x|k) are known.The model uses empirically determined P(x, k) to obtain P(x = 1|k).
- Model scope: The statistical model explains most empirical observations, but networks sharing p(k) and assortativity r can still exhibit different paradox magnitudes.Accurate estimation therefore requires the more detailed joint degree distribution e(k, k′).
III. DISCUSSION
The discussion connects the majority illusion to biased local sampling by highly connected nodes. It argues that disassortative structure and small structural or attribute-correlation changes can increase susceptibility to influence and external shocks.
- Mechanism: High-degree nodes skew many people’s observations because they are overrepresented in local neighborhoods.This links the majority illusion to the friendship paradox and to the influence of highly connected nodes.
- Network structure: The paradox is much stronger in disassortative networks, where high-degree nodes tend to connect to low-degree nodes.For a fixed degree distribution, this gives high-degree nodes greater power to skew others’ observations than in assortative networks.
- Implications: Some network structures may therefore be more susceptible to influence manipulation and the spread of external shocks.The discussion presents this as a susceptibility implication of the measured structural differences.
- Implications: Small changes in topology, assortativity, or degree–attribute correlation may exacerbate the paradox without changing the attribute distribution.The paper relates this possibility to apparently sudden shifts in public attitudes.
- Related concept: The majority illusion is an example of class size bias, in which more popular classes or events are overrepresented in samples.This provides a broader sampling-bias connection for the network phenomenon.
S1. DATA
The study evaluates the majority illusion using synthetic scale-free and Erdős-Rényi-type networks alongside several real-world social or collaboration networks.
- Synthetic networks: Scale-free networks were generated with the configuration model from a specified power-law degree sequence p(k) ∼ k^-α.Random pairs of half-edges were linked until the construction could proceed no further.
- Synthetic networks: Erdős-Rényi-type networks used 10,000 nodes with random links chosen to match the average degrees of the scale-free networks.The linking probability was fixed to produce comparable average degree.
- Real-world networks: Real-world data included the HepTh high-energy-physics collaboration network, the Digg follower graph, and a political-blog network.These networks were summarized using their observed statistics.
A. Friendship Paradox
The friendship paradox arises because neighbors are sampled through edges, which overrepresents high-degree nodes relative to random nodes. Degree heterogeneity and positive degree–attribute correlation extend this bias to attributes such as activity.
- Degree bias: The neighbor degree distribution is q(k) = kp(k)/⟨k⟩, because nodes with more edges are more likely to be reached by a randomly chosen edge.This differs from the node degree distribution p(k).
- Degree bias: Average neighbor degree exceeds average node degree, and the friendship paradox becomes more pronounced as degree heterogeneity σk increases.The difference is nonnegative because σk ≥ 0.
- Attribute bias: For a binary attribute, global activity is weighted by p(k), whereas activity among random neighbors is weighted by the neighbor distribution q(k).Active nodes are designated x = 1 and inactive nodes x = 0.
- Attribute bias: The generalized friendship paradox is positive when degree and attribute are positively correlated, with strength increasing with ρkx.The correlation relates active-node average degree to overall average degree.
B. Gaussian Approximation
For some well-behaved degree distributions, a multivariate-normal approximation estimates the conditional probability needed to calculate majority-illusion strength. It fits the α = 3.1 network well but deviates substantially for heavier-tailed networks.
- B. Gaussian Approximation: The method approximates P(x′, k′) as a multivariate normal distribution to estimate P(x′ = 1|k′).This conditional probability is required to calculate majority-illusion strength using Eq. 4.
- B. Gaussian Approximation: The approximation applies analytically to some well-behaved degree distributions, including scale-free p(k) ∼k−α with α > 3 and near-zero-assortativity Poisson networks.
- B. Gaussian Approximation: The Gaussian approximation fits the α = 3.1 synthetic scale-free network quite well.Figure S1 compares empirical paradox-regime fractions with theoretical estimates from the approximation.
- B. Gaussian Approximation: The theoretical estimate deviates significantly from data for the heavier-tailed α = 2.1 network.The same figure reports empirical fractions as symbols and Gaussian-approximation estimates as dashed lines.
C. Influence of Network Structure
Degree sequence and assortativity do not fully determine network structure, and structural differences can alter majority-illusion strength. Rewired synthetic networks with identical degree sequences and assortativity show some variation, especially at midassortativity.
- C. Influence of Network Structure: Networks sharing a degree sequence and assortativity can still have different structures that affect majority-illusion strength.A network is not fully specified by its degree sequence and degree assortativity r.
- C. Influence of Network Structure: Edge rewiring changes network structure while preserving both the degree sequence and assortativity r.Structural constraints limit the attainable assortativity range for a given degree sequence.
- C. Influence of Network Structure: Figure S2 compares the fraction of nodes in the paradox regime across synthetic scale-free networks with controlled degree sequences and assortativity.Identical symbols within a plot denote networks with the same assortativity.