12,404,868 LiveJournal posts with the mood their authors selected, labelled negative (5,014,041), neutral (2,894,738) or positive (4,496,089), by 115,444 users, with the time of each post; the friendship links (246,164) and follower links (793,948) among the users; and the 2,473,074 memberships of 84,925 users in 200,208 communities. The data were crawled from LiveJournal within three hops of a starting user and preprocessed by Shengmin Jin and Reza Zafarani for their paper 'Emotions in Social Networks: Distributions, Patterns, and Models' (CIKM 2017); users and communities are identified only by numbers assigned by the authors.
Multiplex networks of Twitter users during 5 events of 2013 and 2014 (Cannes2013, MLKing2013, MoscowAthletics2013, NYClimateMarch2014, ObamaInIsrael2013), as collected by Elisa Omodei, Manlio De Domenico and Alex Arenas and distributed by Netzschleuder. A node is a user, identified only by an anonymous number; a directed link from i to j means that user i retweeted, mentioned or replied to user j during the event, weighted by the number of such interactions. The package holds 3,238,746 nodes and 6,014,230 links (2,921,236 retweets, 2,383,605 mentions, 709,389 replies) of 5 networks; the column layer names the event, so a single event is a filter of each table.
The GitHub Stargazers graph classification dataset of the Karate Club project: 12,725 small social networks of developers who starred popular machine-learning or web-development repositories on GitHub (up to August 2019). Nodes are users and links are follower relationships; each graph is the largest connected component of its repository group (at least 10 users) and has a binary label. This copy holds 2,985,781 links and 12,725 graph labels (6,808 labelled 1 and 5,917 labelled 0).
Twitter news-sharing networks built from COVID-related tweets to detect low-credibility accounts, as published by Truong, Allen and Menczer (OSoMe) on Zenodo: a retweet network among users (434,963 edges), a bipartite network of users and the news domains they shared (613,608 edges) and a user-user co-sharing network (48,770,798 edges), 49,819,369 edges in all. Edges carry a weight and the credibility labels and scores of their ends: a label of 0 (high), 1 (low) or -1 (unknown) and, for accounts, a score. Users are Twitter user IDs.
Retweets of tweets that link to low-credibility ('claim') or fact-checking articles around the 2016 US presidential election, from the study 'Anatomy of an online misinformation network' (Shao and colleagues, PLOS ONE 2018), in the two files the archive provides: the file named pre-election has 2,132,235 retweets from 2016-05-16 to 2016-11-07, each labelled claim or fact-checking, and the file named post-election has 5,816,179 claim-only retweets from 2016-05-16 to 2017-10-09. The second file also contains all 1,746,867 claim rows of the first, so the two layers overlap; 7,948,414 rows in all. Each row is one retweet, with the user retweeted, the user retweeting, the tweet identifier and the time.
Eight networks of Facebook pages from the GEMSEC study (Rozemberczki, Davies, Sarkar and Sutton): the pages of one category are the nodes and an edge joins two pages that like each other (mutual likes). 1,380,293 edges in all: artist (819,306 links), athletes (86,858 links), company (52,310 links), government (89,455 links), new sites (206,259 links), politician (41,729 links), public figure (67,114 links), tvshow (17,262 links). The pages are identified by numbers that restart in each category (the column layer names the category).
The Twitch social networks of the MUSAE project: for each of seven language communities of Twitch streamers (DE German, ENGB English, ES Spanish, FR French, PTBR Portuguese, RU Russian, ZHTW Chinese), a network of mutual friendships between streamers, with attributes of each streamer (account age in days, mature-content flag, view count, partner flag) and a list of feature identifiers. This copy holds 492,575 links among 36,890 streamers in all seven networks. SNAP describes the data as 'Twitch user-user networks of gamers who stream in a certain language', collected in May 2018 for a binary classification task (predicting whether a streamer uses explicit language).
Two bipartite edit networks of Wikimedia projects, as the Koblenz Network Collection (KONECT) distributes them: edit-enwiki, the edits of the English Wikipedia (266,769,613 edits by 3,819,691 users on 21,416,395 articles, from 17 January 2001 to 6 September 2010) and edit-frwikinews, the edits of French Wikinews (193,618 edits by 1,408 users on 25,039 articles, from 5 December 2003 to 31 August 2010). Each row is one edit, with the edit's Unix timestamp. The files hold node numbers only, no user names, article titles or text.
Three networks of political communication among Twitter users in the six weeks before the 2010 US Congressional midterm elections, from the study by Conover and colleagues ('Political Polarization on Twitter'): the retweet network (61,157 edges among 18,470 users), the mention network (15,159 edges among 7,175 users) and a combined network of both kinds of edge (77,920 edges among 22,405 users). Each edge is one tweet with its type, number of hyperlinks, time and hashtags, and each user has a cluster label (left, right, or none) from the retweet network.
A multi-dimensional YouTube network crawled in December 2008 by Lei Tang, Xufei Wang and Huan Liu: between 15,088 active users, five kinds of interaction (the contact network, and the numbers of shared friends, subscriptions, subscribers and favourite videos), each with an intensity. This copy holds all five as layers of one table, 13,628,895 rows among 14,992 users, plus the list of 15,088 users. The networks are symmetric and each pair is stored once.
The Facebook Social Connectedness Index (SCI) release of January 2026, published on the Humanitarian Data Exchange by Meta's AI for Good programme: for pairs of geographic units, an index of the relative probability that a Facebook user in one unit is friends with a user in the other. This copy holds 3,085,880,396 rows in 19 layers: pairs of countries, of regions at several levels (GADM levels 1 and 2, geoBoundaries ADM1 and ADM2, European NUTS levels 1 to 3, US counties and US ZIP code tabulation areas) and of a region with a country. The SCI is scaled within each file to lie between 1 and 1,000,000,000.
The rugby dataset of the multi-view Twitter data of Derek Greene and Pádraig Cunningham (2013): 854 Twitter accounts, collected in 2012 from curated Twitter lists. International Rugby Union players, clubs and organisations active on Twitter; the ground-truth communities correspond to 15 countries and overlap, because players can be assigned to both their home nation and the nation in which they play club rugby. The accounts are described by nine views, given upstream as sparse matrices: six relation views between the accounts (who follows whom: 35,757 links; mentions: 33,832 account pairs; retweets: 12,472 pairs; each relation is listed in both directions, as a pair of views) and three feature views (words from the accounts' tweets: 955,307 entries over 35,352 words; Twitter lists the accounts appear on: 77,023 entries over 5,900 lists; merged list names: 74,589 entries over 2,840 words). The package also holds the 15 ground-truth communities (1,051 memberships). Accounts are identified by their Twitter IDs only; the upstream archive contains no tweets and no names.
The olympics dataset of the multi-view Twitter data of Derek Greene and Pádraig Cunningham (2013): 464 Twitter accounts, collected in 2012 from curated Twitter lists. Athletes and organisations involved in the London 2012 Summer Olympics; the ground-truth communities correspond to 28 different sports. The accounts are described by nine views, given upstream as sparse matrices: six relation views between the accounts (who follows whom: 10,642 links; mentions: 9,615 account pairs; retweets: 3,740 pairs; each relation is listed in both directions, as a pair of views) and three feature views (words from the accounts' tweets: 515,393 entries over 18,455 words; Twitter lists the accounts appear on: 46,478 entries over 4,942 lists; merged list names: 64,613 entries over 3,097 words). The package also holds the 28 ground-truth communities (464 memberships). Accounts are identified by their Twitter IDs only; the upstream archive contains no tweets and no names.
The politics-uk dataset of the multi-view Twitter data of Derek Greene and Pádraig Cunningham (2013): 419 Twitter accounts, collected in 2012 from curated Twitter lists. Members of Parliament of the United Kingdom; the ground truth consists of five groups corresponding to political parties. The accounts are described by nine views, given upstream as sparse matrices: six relation views between the accounts (who follows whom: 27,340 links; mentions: 14,788 account pairs; retweets: 7,270 pairs; each relation is listed in both directions, as a pair of views) and three feature views (words from the accounts' tweets: 608,614 entries over 19,868 words; Twitter lists the accounts appear on: 61,064 entries over 3,614 lists; merged list names: 78,364 entries over 2,879 words). The package also holds the 5 ground-truth communities (419 memberships). Accounts are identified by their Twitter IDs only; the upstream archive contains no tweets and no names.
The politics-ie dataset of the multi-view Twitter data of Derek Greene and Pádraig Cunningham (2013): 348 Twitter accounts, collected in 2012 from curated Twitter lists. Irish politicians and political organisations, assigned to seven disjoint ground-truth groups according to their affiliation. The accounts are described by nine views, given upstream as sparse matrices: six relation views between the accounts (who follows whom: 16,856 links; mentions: 6,318 account pairs; retweets: 3,019 pairs; each relation is listed in both directions, as a pair of views) and three feature views (words from the accounts' tweets: 350,175 entries over 14,377 words; Twitter lists the accounts appear on: 15,225 entries over 1,047 lists; merged list names: 25,651 entries over 1,051 words). The package also holds the 7 ground-truth communities (348 memberships). Accounts are identified by their Twitter IDs only; the upstream archive contains no tweets and no names.
The football dataset of the multi-view Twitter data of Derek Greene and Pádraig Cunningham (2013): 248 Twitter accounts, collected in 2012 from curated Twitter lists. A collection of English Premier League football players and clubs active on Twitter; the ground-truth communities correspond to the 20 individual clubs of the league. The accounts are described by nine views, given upstream as sparse matrices: six relation views between the accounts (who follows whom: 3,819 links; mentions: 3,312 account pairs; retweets: 1,350 pairs; each relation is listed in both directions, as a pair of views) and three feature views (words from the accounts' tweets: 224,890 entries over 11,806 words; Twitter lists the accounts appear on: 64,529 entries over 7,814 lists; merged list names: 58,528 entries over 3,601 words). The package also holds the 20 ground-truth communities (248 memberships). Accounts are identified by their Twitter IDs only; the upstream archive contains no tweets and no names.
LiveJournal
Twitter (X)
GitHub
Facebook
Twitch
Wikipedia
YouTube