Planetary-Scale Views on an Instant-Messaging Network

1 Introduction

• The study analyzes one month of anonymized Messenger activity, covering 30 billion conversations generated by 240 million users.The authors did not access message content and focus on collective patterns rather than individual behavior.Source: §1 · p. 2
• The communication graph has 180 million nodes and 1.3 billion undirected edges, with an edge for each pair that exchanged a message.This graph represents observed communication, not buddy-list connections.Source: §1 · p. 2
• 99.9% of nodes belong to the largest connected component; average shortest path length is 6.6 (median 6), with paths up to 29.The clustering coefficient decays with exponent −0.37, lower than expected from prior research (11).Source: §1 · p. 2
• People communicate more and for longer with others similar in language, location, and age, with homophily strongest for language and weakest for age.Gender is an exception: cross-gender conversations are more frequent and longer than same-gender conversations.Source: §1 · p. 2
• Conversation frequency decreases with geographic distance, while links spanning longer distances tend to carry more and longer conversations.Source: §1 · p. 3

2 Instant Messaging

• Instant messaging provides fast, near-synchronous communication using short text messages, while users and contacts are commonly called buddies.Earlier IM studies examined 16 users, a 140,000-person contact network, or 400 users in one organization; this study covers the full Messenger population over one month.Source: §2 · p. 3, §2.1 · p. 3

2.2 Data description

• The study combines demographic information, time-stamped presence events, and communication-session logs, analyzing anonymized high-level patterns rather than individuals or message text.The recorded demographic data include self-reported age, gender, location, and language; IP addresses were used to derive geographic coordinates.Source: §2.2 · pp. 3–4
• Sessions and conversations mean IM interactions among two or more people, and participants may join or leave over time.The system limits simultaneous participation to 20, though the paper observed long sessions with more than 50 participants joining over time.Source: §2.2 · p. 3
• The authors parsed 30 days of logs through a multithreaded pipeline and addressed missing events and recycled session IDs by closing sessions after 48 hours without a leave event.Parsing compressed the logs and converted join/leave events into conversations; sessions were also closed automatically when only one user remained.Source: §2.2 · p. 4

3.1 Levels of activity

• 242,720,596 users logged in during the observation period, and 179,792,538 were active in conversations; on June 1, 982,005,323 sessions exchanged more than 7 billion messages.The section reports aggregated counts across the observation period and a representative day.Source: §3.1 · p. 4
• Logins per user are heavy-tailed with exponent 3.6, while buddy additions spike at 600, the contact-list limit.Figure 1 displays the per-user login and buddy-addition distributions.Source: §3.1 · p. 4, Figure 1 · p. 5
• Conversation participation peaks at 20, the simultaneous limit, and session durations fit a power law with exponent 3.6.Figure 2 presents participant counts and session-duration distributions.Source: §3.1 · p. 4, Figure 2 · p. 5
• Online durations fit exponent 1.77 and offline intervals 1.3; online spells are shorter and decay faster, with periodicities at 12, 24, and 48 hours.Logout intervals also show periodicities at multiples of 24 hours.Source: §3.1 · p. 5, Figure 3 · p. 6
• Figure 4 reports weekday averages for logins, status changes, and buddy additions, using five weeks of data to compute error bars.Login-event counts exceed distinct users logging in because a user can log in multiple times per day.Source: §3.1 · pp. 5–6, Figure 4 · p. 6

3.2 Demographic characteristics of the users

• Compared with 2005 world census data, Messenger users aged 15–35 are strongly overrepresented; females are overrepresented at ages 10–14, and 6.5% did not report age.Figure 5 visualizes the population comparison.Source: §3.2 · pp. 6–7, Figure 5 · p. 7
• Figure 7 plots average conversation duration per user and the time between a user's conversations.Source: Figure 7 · p. 8
• The self-reported age distribution is right-skewed, has a mode at 18, and has exponential tails.Figure 6 also plots relative age-group differences between Messenger and world populations.Source: Figure 6 · p. 7, §3.2 · p. 8

• The analysis covers two-person conversations, which account for 99% of all conversations.Source: §4 · p. 8
• Conversation lengths follow a heavy-tailed distribution with exponent -3.7 and a mode of 4 minutes.Source: §4 · p. 8
• Intervals between consecutive conversations follow a power law with exponent −1.5, which the authors explain using a priority-queue model.The passage notes a similar temporal pattern in email and letter communication.Source: §4 · p. 8
• Conversation and message totals between user pairs are heavy-tailed but do not appear to follow a power law.Figure 8 presents the number of conversations per user in a month and messages exchanged per conversation.Source: §4 · p. 8, Figure 8 · p. 9

5.1 Communication by age

• Older people tend to have longer conversations and exchange more messages per conversation; younger people have faster-paced dialogs, whereas older people exchange messages at a slower pace.The messages-per-conversation plot shows a dip for ages 25–45.Source: §5.1 · pp. 9–10
• Users ages 10–35 show strong age similarity in communication, while users 35 and older communicate more evenly across ages; older users have the longest conversations, with a duration valley at ages 25–35.Source: §5.1 · p. 10

5.2 Communication by gender

• About 50% of conversations are cross-gender and 40% are between users of the same gender; observed tie shares are 25% male–male, 21% female–female, and 54% cross-gender, close to chance expectations.Source: §5.2 · pp. 11–12
• Female–male conversations average more than 5 minutes and 7.6 messages, compared with 4.5 minutes and 6.6 messages for female–female, and about 4 minutes and 5.9 messages for male–male conversations.Source: §5.2 · p. 11

5.3 World geography and communication

• Messenger users are dense in North America, Europe, and Japan, while per-capita usage is high in places including Scandinavia, Ireland, Australia, and South Korea.The geographic map uses login locations; the per-capita view accounts for population density.Source: §5.3 · p. 12, Figure 12 · p. 13
• International communication flows link Australia and New Zealand toward Europe and the United States, while Brazilian communications are weighted toward Europe and Asia.Source: §5.3 · p. 13, Figure 13 · p. 14

5.4 Communication among countries

• The United States and Spain appear as country-communication hubs, while the conversation-duration network’s core appears to be Arabic countries.Edges represent country pairs with more than 10 million monthly conversations; the duration view weights edges by mean conversation length.Source: §5.4 · pp. 14–15, Figure 14 · p. 14
• Northern European countries and Canada rank highly in population share using Messenger, while Afghanistan tops the list for conversations per user and for messages and minutes talked per user per day.Source: §5.4 · pp. 15–16

• Compared with random pairs, communicating pairs show stronger attribute correlations and are more likely to share age, location, or language, but less likely to be the same gender.Source: §6 · p. 18
• Age distributions for communicating pairs show a diagonal trend, indicating that people tend to communicate with others of similar reported age.Source: §6 · p. 19, Figure 16 · p. 19
• People with similar attributes tend to have more frequently used links with shorter, more intense conversations, and communication intensity decays linearly with age difference.Source: §6 · p. 19, Figure 17 · p. 20
• People who communicate are also more likely to search the web for content on similar topics.The authors attribute this related finding to subsequent work using their data.Source: §6 · p. 19

7 The communication network

• The communication network has a heavy-tailed degree distribution, but it does not follow a power law.A power law with exponential cutoff fit has parameters a = 0.8 and b = 0.03.Source: §7 · p. 21
• 99.9% of nodes belong to the largest connected component, and the average clustering coefficient is 0.137.The clustering coefficient decays slowly with degree, with exponent −0.37.Source: Figure 19 · p. 22, §7 · p. 22

7.1 How small is the small world?

• 6.6 hops is the average shortest path length; the mode is 6 hops and the 90% effective diameter is 7.8.The estimate used 1,000 randomly sampled nodes, with paths calculated from each to all other nodes.Source: Figure 20 · p. 22, §7.1 · p. 23

7.2 Network cores

• 79 nodes form the central core, each with more than 68 edges within the set.Core sizes remain stable until k ≈ 20, then drop rapidly.Source: §7.2 · p. 24

7.3 Strength of the ties

• Removing high-link, conversation, total-time, or sent-message users shrinks the largest component rapidly; random removal shrinks it slowest.After 160 million of 180 million nodes are removed randomly, the largest component still contains about half the nodes.Source: Figure 22 · p. 25, §7.3 · p. 25
• Removing users by link count, conversations, total conversation time, or sent messages removes edges fastest; average messages per conversation removes them slowest.The authors suggest users with longer conversations and more messages per conversation tend to have smaller degrees.Source: §7.3 · pp. 25–26, Figure 23 · p. 26

8 Conclusion

• The dataset covers more than 30 billion conversations among 240 million people, and its communication graph has 180 million nodes and 1.3 billion edges.The graph is described as the largest social network analyzed to date.Source: §8 · pp. 26–27
• The graph is well connected, highly transitive, and robust; similar age, language, and location are associated with more communication, while cross-gender conversations are more frequent and longer.The paper also reports planetary-scale validation of the earlier “6 degrees of separation” finding.Source: §8 · p. 27
• The dataset’s scale limits the analyses that can be performed; sampling may corrupt network properties, and full data are needed to measure age and distance homophily reliably.The authors identify degree distribution and graph diameter as properties that sampling can corrupt.Source: §8 · p. 27
• Future research directions include predictive models of communication patterns and studying network evolution over time.The models use structural and demographic attributes to forecast communication frequencies and conversation durations.Source: §8 · p. 27