Source-linked AI summary
Calling Dunbar's Numbers
Pádraig MacCarron, Kimmo Kaski, Robin Dunbar
TL;DR
This paper asks whether social layers predicted by the social brain hypothesis can be detected from mobile-phone call frequencies and whether their sizes match prior estimates. It normalizes each user's call frequencies and applies clustering algorithms; the analyses find evidence for layered structure, with stronger agreement for outer layers and variability among methods.
Problem
The paper asks whether layers of friends can be detected from phone-call frequencies in an offline context and whether their sizes match previously identified layers.
Source: §I · p. 1Method
The study normalizes each ego's alter call frequencies and applies clustering algorithms to compare the resulting layer sizes with social-brain predictions.
Source: §I · p. 1, §II · p. 2Results
The methods find a small number of clusters and support the outer two layers; Jenks and Gaussian Mixture Model results for four layers are close to Dunbar layers, though methods differ and lower-usage users match the inner three.
Source: §IV · pp. 5–7Takeaways & Limitations
Similar social-network structure across communication modes has implications for the social brain hypothesis, while the paper proposes future study of layer turnover over time.
Source: §IV · p. 7Finding discrete layers remains a considerable challenge, despite evidence of structure in the noisy dataset.
Source: §IV · p. 5
Abstract
from arXiv · showhide
The social brain hypothesis predicts that humans have an average of about 150 relationships at any given time. Within this 150, there are layers of friends of an ego, where the number of friends in a layer increases as the emotional closeness decreases. Here we analyse a mobile phone dataset, firstly, to ascertain whether layers of friends can be identified based on call frequency. We then apply different clustering algorithms to break the call frequency of egos into clusters and compare the number of alters in each cluster with the layer size predicted by the social brain hypothesis. In this dataset we find strong evidence for the existence of a layered structure. The clustering yields results that match well with previous studies for the innermost and outermost layers, but for layers in between we observe large variability.
I. INTRODUCTION
The social brain hypothesis proposes that ego networks form layers of increasing size as emotional closeness decreases, with about 150 friends across the cumulative layers. This study tests whether such layers can be detected from phone calls and whether their sizes match prior estimates.
I. INTRODUCTION
- The social brain hypothesis links primate brain evolution to maintaining larger social groups and predicts ego-network layers that grow as emotional closeness declines.Cumulative layer sizes are reported as about 5, 15, 50, and 150 friends.Source: §I · p. 1
- Prior work has observed Dunbar layers in online social media and an online computer game, while the 150-friend figure describes relationships at a given time.The cited work describes outer layers as changing faster than inner ones.Source: §I · p. 1
- The study tests whether layers of friends can be detected from mobile phone calls and whether their sizes match those previously identified.The analysis uses call frequency to examine layers in an offline context.Source: §I · p. 1
- The dataset contains 34.9 million users and almost 6 billion calls across 2007, with complete call records for about 6 million company users.The company provides coverage to approximately 20% of the country's population.Source: §I · p. 1
- Call frequency is used as a relationship-strength measure because it has been shown to correlate with emotional closeness.Social signatures in cellphone data have also been reported to remain robust over time despite alter identity changes.Source: §I · p. 1
II. METHODS
The methods use phone-call frequencies as one-dimensional measures of ego–alter relationships, normalize these values within each ego, and apply clustering and distribution-fitting procedures. User-selection thresholds address variation in phone use and possible under-reporting.
II. METHODS
- Reciprocal calling is required to exclude casual and business calls, and users are selected by alter-count cutoffs to reduce under-reporting from less frequent phone use.The analysis begins with a 100-alter cutoff, then lowers it to 50 to examine lower-frequency users.Source: §II · pp. 1–2
- Maximum likelihood estimates compare power-law, exponential, stretched-exponential, Gaussian, and log-normal degree distributions, with Akaike Information Criteria selecting the best model.The analysis distinguishes degree, the number of alters called, from weighted degree, the total number of calls.Source: §II · p. 2
- Each ego's alter call counts are normalized by that ego's minimum and maximum counts so the weakest relationship is 0 and the strongest is 1.The study examines the weight distribution and uses local minima in its kernel density estimate to identify clusters.Source: §II · p. 2
- Jenks natural breaks clusters one-dimensional values by minimizing within-cluster distances and maximizing separation between cluster centers; a goodness-of-fit threshold of 0.85 selects cluster count.A goodness of fit of 1.0 requires zero within-class variation.Source: §II · p. 2
- The Gaussian Mixture Model treats the data as generated by several Gaussian distributions and uses the Akaike Information Criterion to assess cluster count.The study also considers a model with means at the Dunbar numbers.Source: §II · p. 2
- Head/tail breaks recursively splits values above the mean, a procedure developed for heavy-tailed data whose cluster count emerges from the distribution.The paper describes its data as heavy tailed, with many users calling many people infrequently and closer friends frequently.Source: §II · pp. 2–3
III. RESULTS
Clustering call frequencies supports a layered social structure, but layer sizes vary substantially. Jenks and Gaussian mixture results align most clearly with the outer two layers.
III. RESULTS
- Log-normal models had slightly more support for degree distributions, while log-normal was best supported for weighted degree among the candidate models.The local-minima analysis identified average groups of 15 alters up to normalized relationship strength 0.28, then an average group of 68.8; beyond that, the data were too noisy to split further at group level.Source: §III · p. 3
- Four clusters were most common, occurring for 7,226 users (27.1%), with average cumulative layer sizes of 4.1, 11.0, 29.8, and 128.9 users, slightly smaller than conventional Dunbar-layer numbers.The next most common was five clusters for 5,983 users (22.4%), with cumulative layer sizes of 2.9, 7.4, 17.7, 43.0, and 134.3 users.Source: §III · p. 3
- For Gaussian mixtures, 9,289 egos formed two clusters (43.0, 122.9 alters), while 3,334 formed three (18.0, 59.7, 126.4), missing the innermost layer.Source: §III · p. 4
- For users with more than 100 alters, head/tail breaks found 12,951 users in three clusters (cumulative layer sizes 4.3, 22.4, 122.9) and 11,616 users in four (2.2, 8.2, 29.8, 133.5).Source: §III · pp. 4–5
- All three algorithms find few layers for most egos; Jenks and Gaussian mixtures support the outer two layers, while layer-size distributions show substantial variation.For 4- and 5-cluster results, alter counts vary widely around their means, extending far to the right.Source: §III · pp. 3–5
- With everyone forced into four Jenks layers, mean sizes were 3.5, 10.6, 31.1, and 129.9 alters; the scaling ratio was 3.3, versus 3.2 in Zhou et al. (2005) and Dunbar et al. (2015).Source: §III · p. 5
- Call duration and frequency correlate at r = 0.71 overall (r = 0.76 for egos with k ≥100); Jenks gives four duration-based layers averaging 3.9, 10.1, 27.2, and 129.3 alters.Source: §III · p. 5
- Among egos with 50 ≤k < 100, three Jenks clusters were most common (23.5%), with cumulative sizes of 3.9, 11.9, and 63.91 alters; duration results were similar.Four clusters were next most common (14.5%), with sizes 2.7, 6.7, 17.9, and 64.3 alters.Source: §III · p. 5
IV. CONCLUSIONS
Different clustering methods support a layered structure in phone-call networks, particularly the outer two layers, although discrete layers remain difficult to identify. Four-layer results are close to Dunbar layer sizes and show a similar scaling pattern, while cluster sizes are right-skewed and variable.
IV. CONCLUSIONS
- All clustering methods find few clusters and support the outer two layers, despite noisy data.The authors note that identifying discrete layers remains a considerable challenge.Source: §IV · p. 5
- Four-layer results are close to Dunbar layer sizes and show a scaling pattern of about 3, consistent with earlier reports across social networks, online platforms, and gaming environments.The authors note that no principled explanation for this consistent pattern is yet available.Source: §IV · p. 6
- Cluster means are generally smaller than predicted Dunbar layer means, though they match the ranges reported by Hill and Dunbar (2003).The authors suggest phone data captures only part of an ego’s social network, even for users who call more than 100 alters.Source: §IV · p. 6
- Users who call 50 to 100 people throughout the year match the inner three Dunbar layers well.Source: §IV · p. 7
- Layer sizes are log-normally distributed across algorithms and cluster counts, with rightward skew and large standard deviations.Source: §IV · p. 7
- Despite the skewed degree distribution, users with many more friends than the mean still show evidence of layers.The authors suggest this pattern may distinguish extroverts.Source: §IV · p. 7