Source-linked AI summary
Validation of Dunbar's number in Twitter conversations
Bruno Goncalves, Nicola Perra, Alessandro Vespignani
TL;DR
The paper asks whether online communication can exceed the cognitive limit on stable social relationships known as Dunbar’s number. It analyzes six months of Twitter conversations and models user behavior with finite priority queues and time resources, finding a maximum of 100–200 stable relationships and behavior reproduced by the model.
Problem
The paper examines whether online communication changes the cognitive limit on the number of stable social relationships known as Dunbar’s number.
Method
The study analyzes Twitter conversations and models users with finite message queues, time resources, and priorities proportional to senders’ total degree.
Results
100–200 stable relationships were observed, and the model reproduced the observed social behavior.
Takeaways & Limitations
Online social networks do not overcome the cognitive and biological constraints limiting stable social relations.
Abstract
from arXiv · showhide
Modern society's increasing dependency on online tools for both work and recreation opens up unique opportunities for the study of social interactions. A large survey of online exchanges or conversations on Twitter, collected across six months involving 1.7 million individuals is presented here. We test the theoretical cognitive limit on the number of stable social relationships known as Dunbar's number. We find that users can entertain a maximum of 100-200 stable relationships in support for Dunbar's prediction. The "economy of attention" is limited in the online world by cognitive and biological constraints as predicted by Dunbar's theory. Inspired by this empirical evidence we propose a simple dynamical mechanism, based on finite priority queuing and time resources, that reproduces the observed social behavior.
I. INTRODUCTION
Online communication tools provide unprecedented data for studying social behavior while raising questions about how digital connectivity affects social relations under finite attention and time.
- Online tools have generated unprecedented data on social behavior and enabled computational approaches to long-standing social-science questions.
- Continuous digital communication appears to expand opportunities for establishing and maintaining relations, blurring distinctions between physical and online social realities.
- 291 typical group members were measured using two methods, roughly twice Dunbar’s estimate, highlighting variation in estimates of social group size.
- Finite time and attention require people to prioritize among competing preferences, interests, and needs when managing social interactions.
II. THE DATASET
The study combines Twitter firehose and API data to assemble a large, multi-year activity record for analyzing conversations and social relations.
- Over six months, the researchers mined Twitter’s firehose to identify a large sample of active accounts.
- 3 million users’ complete histories were queried through the Twitter API.
- Over 380 million tweets covered almost four years of user activity in the resulting dataset.
- The dataset focuses on Twitter conversations accrued over six months to investigate stable social relations mediated by the platform.
A. Tree Identification and Projection
The paper identifies reply conversations as tweet trees, extracts connected components, and projects them into a weighted directed user network representing interaction strength.
- Each reply tweet links to its original tweet and therefore maps directly to a directed edge.
- Depth-first search identifies connected components, treating links as undirected so complete trees can be recovered from any starting node.
- A user graph connects A to B when A replies to B, with repeated replies represented by the directed edge’s weight.
B. Online conversations
Twitter conversations are represented as diverse reply trees and direct user interactions, with network degree based on actual replies rather than follower counts.
- Over 25 million conversation trees were extracted, ranging from small exchanges to conversations with thousands of tweets and hundreds of users.
- Directed user-user networks are obtained by projecting conversation trees to represent how users interact and establish relationships.
- Tree size, shell count, user count, and depth have broad-tailed distributions, indicating diverse user behaviors.
- Mutual interactions produce bidirectional edges, while stronger weights indicate more frequent or prolonged interaction.
- Out degree counts users replied to, excluding followers or friends without direct contact.
III. DUNBAR’S NUMBER IN OUR DATA
The analysis defines stable social relationships through direct reply activity and finds that interaction strength peaks around 100–200 contacts, while reciprocation saturates as incoming connections grow.
- Out-degree counts the distinct users a person replies to, excluding followers without direct interaction.
- The study measures average weight per outgoing edge over a time window as the strength of active relationships.Higher edge weight represents stronger interaction, while the time window controls data aggregation.
- 100–200 friends marks the maximum of average interaction strength, agreeing with Dunbar’s prediction.The result suggests online networks do not overcome biological and physical limits on stable relationships.
- 200–300 incoming connections marks saturation in reciprocated connections despite continued growth in incoming degree.Beyond this point, additional connections are reciprocated at a much smaller rate and may reflect spurious exchanges.
IV. THE MODEL
The model represents users as nodes in a weighted directed network and simulates message handling with finite queues, prioritized replies, and bounded resources.
- Each user is a node whose out-degree counts replied-to users and whose in-degree counts users sending replies.
- Messages increase directed edge weights, so total outgoing messages quantify each user’s communication activity.
- Finite queue capacity qmax,i forces users to prioritize which incoming messages receive replies.Message priority is proportional to the sender’s total degree.
- At each time step, agents reply to selected queued messages, delete handled messages, add prioritized arrivals, and discard excess messages.
- Broadcast messages initialize endogenous random activity but do not contribute to directed interaction weights.
- The model examines how queue size affects simulated activity curves on a heterogeneous network.The network uses N = 10^5 nodes, γ = −2.4, σ = 10, and runs for T = 2 × 10^4 time steps.
A. Effect of the time window T
The time window T controls message circulation in the model: longer windows raise the observable and slightly shift the activity peak.
- T regulates the maximum number of messages circulating through the network.
- T = 10^4, 1.5 × 10^4, and 2 × 10^4 produce a vertical shift and a small change in peak position.
- Longer windows increase message volume linearly and reduce fluctuations, making the peak clearer and more defined.
B. Effect of broadcast probability p
Broadcast probability p primarily affects users with few contacts, whereas activity becomes independent of p once contact numbers exceed queue capacity.
- Broadcast probability p increases received messages and activity for users with small numbers of contacts.
- When saturation occurs, ωout becomes completely independent of p.
- For contact numbers much larger than queue size, ωout is independent of p, while small-contact activity scales linearly with p.Simulations show p-dependence at small kout and the same behavior across larger kout values.
C. Effect of network’s properties
The model’s results are robust to the scale-free network exponent γ: the peak’s shape and position remain consistent, while its height changes with γ.
- C. Effect of network’s properties: The peak’s shape and position are the same across γ = −2.2, −2.4, −2.6, and −2.8.For smaller γ values, gaps in kout emerge because of the network structure.
- C. Effect of network’s properties: The peak height increases as γ decreases.The authors attribute this to degree redistribution and a stronger selection effect for smaller γ.
- C. Effect of network’s properties: The results are robust on γ despite network-structure differences.
D. Single user: analytical approach
A mean-field analysis of one user explains two regimes in the reply dynamics: activity first increases with outgoing connections, then decreases under queue saturation. Their crossover produces a queue-size-dependent maximum, consistent with simulations.
- D. Single user: analytical approach: The single-user analysis models a user with degree k_i and queue capacity q_max,i receiving messages from neighbors over T time steps.Neighbor priorities are sampled from a distribution, and the model’s communication rules are applied dynamically.
- D. Single user: analytical approach: The number of received messages scales as the square of degree because priority is proportional to degree and incoming connections.
- D. Single user: analytical approach: For k_i ≪ q_max,i, the queue remains unsaturated and the user can reply to all received messages.The received-message count R_t tracks messages arriving at time t, while replies are modeled through successive random fractions of the queue.
- D. Single user: analytical approach: In the unsaturated regime, the average number of replies per connection increases linearly with k_out,i.This linear increase is reported as the behavior of the low-contact regime.
- D. Single user: analytical approach: When contacts exceed queue size, saturation limits replies to a small, priority-selected fraction of incoming messages.The number of replies becomes a random fraction of queue size, with ξ_t uniformly distributed between 0 and 1.
- D. Single user: analytical approach: The two regimes yield a linear increase and a decreasing behavior, with a maximum between them whose position depends on queue size.The analytical arguments agree with the numerical results, including the peak predicted between the opposite regimes.
V. CONCLUSIONS
Online social networks did not expand human capacity for stable relationships beyond Dunbar’s predicted limit. Twitter data and a finite-resource model support cognitive and biological constraints on social activity.
- Twitter conversations support Dunbar’s hypothesis that biological limits constrain the number of stable social relationships.
- The analysis covers six months of Twitter conversations involving millions of individuals.
- A simple model combining finite priority queuing and time resources reproduces the observed social behavior.