Ask Ubuntu is a question-and-answer site on the Stack Exchange network. SNAP builds temporal networks from the Stack Exchange data dump in which each interaction is a directed link (u, v, t): user u answered user v's question (a2q), commented on user v's question (c2q) or commented on user v's answer (c2a) at time t. Node numbers are the `OwnerUserId` of the dump, and t is a Unix timestamp in seconds. This copy holds only the answer-to-question network (sx-askubuntu-a2q): 280,102 answers among 137,517 users, posted between 8 January 2009 and 6 March 2016.
Epinions.com was a consumer review site on which members could mark other members as trusted, forming a Web of Trust that helps determine which reviews users see. This is the signed, directed social network SNAP publishes as soc-sign-epinions: 841,372 links among 131,828 members, each carrying a sign of +1 (717,667 links) or -1 (123,705 links).
A subset of LiveJournal users (those within three hops of one starting user) and the relations among them, prepared by Shengmin Jin and Reza Zafarani for a study of emotions in social networks. Users were anonymised, and only users with at least 10 posts whose mood came from LiveJournal's predefined list were kept. This copy holds the two relation files of the dataset: 246,164 friendship rows and 793,948 follower rows, 1,040,112 rows among 106,577 users.
A sample of the Twitter retweet graph from the study 'Characterizing and Detecting Hateful Users on Twitter' (Horta Ribeiro and colleagues, ICWSM 2018): 2,286,592 directed retweet edges among 100,386 anonymised users, a feature vector of 320 numbers for each user, and a label per user: 544 users annotated as hateful, 4,427 as normal and 95,415 not annotated ('other'). The repository describes the data as a network of 100k users of whom about 5k were annotated as hateful or not, collected through a random walk.
BlogCatalog is a social blog directory website that manages bloggers and their blogs. This is the friendship network among bloggers crawled in June 2010, as distributed by the Data Lab's Social Computing Data Repository (version 2). This copy holds 2,043,701 friendship links among 97,884 users and a list of 97,884 user numbers.
Prosper loans network, a network of the Koblenz Network Collection (KONECT) (Person–person loans). KONECT's file for it has the header `% asym positive`, i.e. a directed network with weights marked `positive`. This copy holds 3,394,979 links among 89,269 nodes, timed from 15 November 2005 to 27 September 2011.
The friendship network of BlogCatalog, a social blog directory, crawled in July 2009 and published by Nitin Agarwal, Xufei Wang and Huan Liu in the Social Computing Data Repository: 88,784 bloggers, identified only by number, and 4,186,390 directed friendship pairs. A row (a, b) means that blogger a is a friend of blogger b.
The Slashdot Zoo signed social network in three snapshots (6 November 2008, 16 February 2009 and 21 February 2009), as SNAP publishes it: users of the technology news site Slashdot tag one another as friends or foes, which are the positive and negative links. This copy holds 1,611,448 links in all, one row per link and snapshot: 2008-11-06: 516,575 links among 77,350 users; 2009-02-16: 545,671 links among 81,867 users; 2009-02-21: 549,202 links among 82,140 users.
Twitter 'circles' (lists) data crawled from public sources, as SNAP publishes them: 973 ego networks combined into one edge list. This copy holds the combined edge list of all the ego networks: 2,420,766 links among 81,306 nodes. The archive of the individual ego networks with their features and circles (twitter.tar.gz) is kept in original/ and is not converted.
A directed, signed network of hyperlinks between Reddit communities, as published by Srijan Kumar, William L. Hamilton, Jure Leskovec and Dan Jurafsky: each of the 571,927 rows is a post in a source subreddit that contains a hyperlink in the title to a post in a target subreddit, among 54,075 subreddits, with the time of the post, the post's ID, a label of -1 (the post is explicitly negative towards the target post; 61,140 rows) or +1 (neutral or positive; 510,787 rows), and the 86 numbers that describe the source post's text. The rows are dated from 31 December 2013 to 30 April 2017.
Slashdot threads network, a network of the Koblenz Network Collection (KONECT) (User–user replies). KONECT's file for it has the header `% asym positive`, i.e. a directed network with weights marked `positive`. This copy holds 140,778 links among 51,083 nodes, timed from 1 January 2004 to 4 September 2006 (apart from one row with timestamp 0).
The reply network of the Higgs Twitter dataset: Twitter activity around the announcement of a new particle with the features of the Higgs boson on 4 July 2012, covering messages posted from 1 to 7 July 2012. A directed, weighted link means that one account replied to another. This copy holds 32,523 reply links among 38,918 accounts.
Networked dialogues on the social network site MySpace, as published on Dryad by Šuvakov, Mitrović, Gligorijević and Tadić: 80,749 messages, each from one user to another, among 36,459 anonymised users, collected over a two-month window. Each message has a time stamp and two emotion scores from the SentiStrength classifier, a positive and a negative one; the message texts are not part of the data.
A directed, signed network of hyperlinks between Reddit communities, as published by Srijan Kumar, William L. Hamilton, Jure Leskovec and Dan Jurafsky: each of the 286,561 rows is a post in a source subreddit that contains a hyperlink in the body to a post in a target subreddit, among 35,776 subreddits, with the time of the post, the post's ID, a label of -1 (the post is explicitly negative towards the target post; 21,070 rows) or +1 (neutral or positive; 265,491 rows), and the 86 numbers that describe the source post's text. The rows are dated from 31 December 2013 to 30 April 2017.
Linux kernel mailing list replies network, a network of the Koblenz Network Collection (KONECT) (Person–person replies). KONECT's file for it has the header `% asym positive`, i.e. a directed network with weights marked `positive`. This copy holds 1,096,440 links among 27,927 nodes, timed from 1 January 2006 to 31 December 2013.
Data files of the paper 'The spread of low-credibility content by social bots' (Shao, Ciampaglia, Varol, Yang, Flammini and Menczer; Nature Communications 2018), a study of the spread of links to low-credibility websites on Twitter by automated accounts (social bots). The package holds the 629,514 retweet links between 25,327 accounts with each account's bot score, the bot scores of 105,124 accounts and of 168,744 accounts that shared mainstream news, centrality measures of 346,573 accounts, activity measures of 11,190 accounts, counts of tweet types for 381,985 items, and the 24 most-shared source websites (18 low credibility, 4 fact checking, 2 satire) with their tweet counts. This is a part of the files of the Zenodo record; the record holds many more, including the tweets.
A temporal network of interactions on Mathematics Overflow, a question-and-answer site on the Stack Exchange network. SNAP builds it from the Stack Exchange data dump: each interaction is a directed link (u, v, t) where user u answered (a2q), or commented on (c2q) a question of user v, or commented on an answer of user v (c2a), at time t. Node numbers are the OwnerUserId of the dump and t is a Unix timestamp in seconds. The package holds the union that SNAP publishes as sx-mathoverflow: 506,550 interactions among 24,818 users (107,581 answers to questions, 203,639 comments on questions, 195,330 comments on answers), made between 29 September 2009 and 6 March 2016. The column edge_type names the kind, so one of the three networks is a filter of the table.
A temporal network of answer to question interactions on Mathematics Overflow, a question-and-answer site on the Stack Exchange network, built by SNAP from the Stack Exchange data dump (sx-mathoverflow-a2q). Each row is a directed link (u, v, t): user u answered the question of user v at time t. The table holds 107,581 interactions among 21,688 users, made between 29 September 2009 and 6 March 2016.
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 temporal network of comment to question interactions on Mathematics Overflow, a question-and-answer site on the Stack Exchange network, built by SNAP from the Stack Exchange data dump (sx-mathoverflow-c2q). Each row is a directed link (u, v, t): user u commented on the question of user v at time t. The table holds 203,639 interactions among 16,836 users, made between 30 September 2009 and 6 March 2016.
A temporal network of comment to answer interactions on Mathematics Overflow, a question-and-answer site on the Stack Exchange network, built by SNAP from the Stack Exchange data dump (sx-mathoverflow-c2a). Each row is a directed link (u, v, t): user u commented on the answer of user v at time t. The table holds 195,330 interactions among 13,840 users, made between 29 September 2009 and 6 March 2016.
Anybeat was an online community that Network Repository describes as a public gathering place where people interact with others from their neighbourhood or across the world. This is the network of links between its members as Network Repository distributes it, as soc-anybeat, in a two-column edge list. This copy holds 67,053 links among 12,645 members.
The Mastodon active-user graph of the Fediverse as crawled by the Fedivertex project: 24 snapshots taken between 5 February 2025 and 15 July 2025. In the upstream documentation this is an active-user graph: for each instance the 10,000 most recently active users are considered; for each of them, the accounts they follow are listed and 1 is added to the edge from X to Y, where X is the instance of the user and Y the instance of the followed account, so the weight of the edge from X to Y encodes how much of the content seen on X is generated on Y; the graph is directed and contains self-loops. The package holds 12,811,937 links, each with the snapshot it was seen in, and 160,468 instance records (8,135 distinct instances; one record per instance per snapshot) with the attributes upstream recorded (`version`, `users`, `statuses`, `languages`, `registration_enabled`). A snapshot lists between 6,185 and 7,142 instances and between 490,701 and 560,299 links. Edge weights are whole numbers from 1 to 548,292 (15,034 distinct values). Instance names are written with "[DOT]" in place of each dot, exactly as upstream published them. The graph is declared directed. Fedivertex covers 13 such graph series (seven Fediverse software systems and several kinds of relation); this package is the Mastodon active-user series only.
Actions of students on a MOOC platform, as published by SNAP for studies of temporal interaction networks: a directed, temporal network from students to course activities. Each of the 411,749 rows is one action of one of 7,047 students on one of 97 targets (course activities), with the number of seconds since the start of the data, a 4-dimensional feature vector per action (411,749 rows) and a label per action (411,749 rows) that is 1 for the 4,066 actions after which the student dropped out.
Trust certifications among members of Advogato, an online community of free-software developers launched in 1999, where members certified one another at one of four levels: Observer, Apprentice, Journeyer or Master. This copy holds 51,332 certifications among 6,551 members. Each is a directed link from the member who gave the certification to the member who received it, weighted 0.6 (Apprentice), 0.8 (Journeyer) or 1.0 (Master); Observer certifications, which express no trust, do not appear. The data were collected by the Trustlet project (Massa, Salvetti and Tomasoni) from the certification graph that Advogato published, and this file is the one Network Repository distributes as soc-advogato, in KONECT's file format. It includes 3,995 self-certifications, and 15 certifications are listed twice.
A who-trusts-whom network of people trading Bitcoin on the Bitcoin OTC platform, collected by Srijan Kumar, Francesca Spezzano, V. S. Subrahmanian and Christos Faloutsos and distributed by SNAP. Each of the 35,592 rows is a directed rating: a member (the source) rated another member (the target) from -10 (total distrust) to +10 (total trust) in whole steps, at a time given in seconds since the Unix epoch. The 35,592 ratings are among 5,881 members, who are identified only by numbers; they were made between 8 November 2010 and 25 January 2016. 32,029 ratings are positive (90%), 3,563 negative.
The Wikispeedia data of Robert West and Jure Leskovec ('Human Wayfinding in Information Networks', WWW 2012), distributed by SNAP: a condensed version of Wikipedia with 4,604 articles and 119,882 hyperlinks between them, the articles' text, a category for the articles (5,204 article-category pairs in 129 categories), and 51,318 finished and 24,875 unfinished navigation paths of players of the online game Wikispeedia, in which a player is given a start and a target article and must click from one to the other using only links. Paths were played between 18 August 2008 and 15 January 2014; each carries a hashed IP address of the player, the time, the duration, the sequence of articles (with back-clicks), and a difficulty rating for finished paths or the target and the way the game ended for unfinished ones.
A who-trusts-whom network of people trading Bitcoin on the Bitcoin Alpha platform, collected by Srijan Kumar, Francesca Spezzano, V. S. Subrahmanian and Christos Faloutsos and distributed by SNAP. Each of the 24,186 rows is a directed rating: a member (the source) rated another member (the target) from -10 (total distrust) to +10 (total trust) in whole steps, at a time given in seconds since the Unix epoch. The 24,186 ratings are among 3,783 members, who are identified only by numbers; they were made between 8 November 2010 and 22 January 2016. 22,650 ratings are positive (94%), 1,536 negative.
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 PeerTube follow graph of the Fediverse as crawled by the Fedivertex project: 62 snapshots taken between 3 February 2025 and 13 April 2026. In the upstream documentation this is a follow graph: an instance X can follow an instance Y to let its users see all the videos posted on Y; the graph is directed, with edges of weight 1 for following. The package holds 1,038,682 links, each with the snapshot it was seen in, and 64,677 instance records (2,356 distinct instances; one record per instance per snapshot) with the attributes upstream recorded (`totalUsers`, `totalDailyActiveUsers`, `totalWeeklyActiveUsers`, `totalMonthlyActiveUsers`, `totalLocalVideos`, `totalVideos`, `totalLocalPlaylists`, `totalVideoComments`, `totalLocalVideoComments`, `totalLocalVideoViews`, `serverVersion`, `totalInstanceFollowers`, `totalPeertubeInstanceFollowers`, `totalInstanceFollowing`, `totalPeertubeInstanceFollowing`, `categories`, `languages`, `error`). A snapshot lists between 836 and 1,756 instances and between 10,067 and 49,104 links. Every link has weight 1. Instance names are written with "[DOT]" in place of each dot, exactly as upstream published them. The graph is declared directed. Fedivertex covers 13 such graph series (seven Fediverse software systems and several kinds of relation); this package is the PeerTube follow series only.
UC Irvine messages network, a network of the Koblenz Network Collection (KONECT) (User–user message). KONECT's file for it has the header `% asym positive`, i.e. a directed network with weights marked `positive`. This copy holds 59,835 links among 1,899 nodes, timed from 15 April 2004 to 25 October 2004.
The Misskey active-user graph of the Fediverse as crawled by the Fedivertex project: 24 snapshots taken between 3 February 2025 and 15 July 2025. In the upstream documentation this is an active-user graph: for each instance the 10,000 most recently active users are considered; for each of them, the accounts they follow are listed and 1 is added to the edge from X to Y, where X is the instance of the user and Y the instance of the followed account, so the weight of the edge from X to Y encodes how much of the content seen on X is generated on Y; the graph is directed and contains self-loops. The package holds 274,652 links, each with the snapshot it was seen in, and 21,425 instance records (1,274 distinct instances; one record per instance per snapshot) with the attributes upstream recorded (`users_count`, `posts_count`, `languages`, `description_language`). A snapshot lists between 841 and 982 instances and between 9,999 and 12,894 links. Edge weights are whole numbers from 1 to 872,529 (2,619 distinct values). Instance names are written with "[DOT]" in place of each dot, exactly as upstream published them. The graph is declared directed. Fedivertex covers 13 such graph series (seven Fediverse software systems and several kinds of relation); this package is the Misskey active-user series only.
Blogs network, a network of the Koblenz Network Collection (KONECT) (Blog–blog hyperlinks). KONECT's file for it has the header `% asym unweighted`, i.e. a directed network with weights marked `unweighted`. This copy holds 19,025 links among 1,224 nodes.
Anonymised email between the members of a large European research institution, as published by SNAP. The package holds two networks that SNAP distributes separately: the static email-Eu-core network (25,571 links among 1,005 members; a link means that one member sent at least one email to another), with the department of each of its 1,005 members, and the email-Eu-core temporal network (332,334 timed emails among 986 members, one row per recipient, spanning 803.9 days). Only communication among members is included, not messages to or from outsiders.
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 Pleroma active-user graph of the Fediverse as crawled by the Fedivertex project: 24 snapshots taken between 3 February 2025 and 15 July 2025. In the upstream documentation this is an active-user graph: for each instance the 10,000 most recently active users are considered; for each of them, the accounts they follow are listed and 1 is added to the edge from X to Y, where X is the instance of the user and Y the instance of the followed account, so the weight of the edge from X to Y encodes how much of the content seen on X is generated on Y; the graph is directed and contains self-loops. The package holds 30,207 links, each with the snapshot it was seen in, and 14,538 instance records (821 distinct instances; one record per instance per snapshot) with the attributes upstream recorded (`version`, `users`, `statuses`, `languages`, `registration_enabled`). A snapshot lists between 540 and 709 instances and between 1,161 and 1,365 links. Edge weights are whole numbers from 1 to 14,259 (209 distinct values). Instance names are written with "[DOT]" in place of each dot, exactly as upstream published them. The graph is declared directed. Fedivertex covers 13 such graph series (seven Fediverse software systems and several kinds of relation); this package is the Pleroma active-user series only.
The Lemmy federation-with-blocks graph of the Fediverse as crawled by the Fedivertex project: 62 snapshots taken between 3 February 2025 and 13 April 2026. In the upstream documentation this is a federation graph completed with negative edges when instance X blocks instance Y; the graph is directed. The package holds 5,641,052 links, each with the snapshot it was seen in, and 22,621 instance records (779 distinct instances; one record per instance per snapshot) with the attributes upstream recorded (`version`, `users`, `posts`, `comments`, `communities`, `users_active_day`, `users_active_week`, `users_active_month`, `users_active_half_year`, `captcha_enabled`, `require_email_verification`, `description_language`, `error`). A snapshot lists between 291 and 439 instances and between 54,200 and 141,669 links. Edge weights are +1 (5,620,855 links) or -1 (20,197 links, the blocks), and the `sign` column repeats them as integers. Instance names are written with "[DOT]" in place of each dot, exactly as upstream published them. The graph is declared directed. Fedivertex covers 13 such graph series (seven Fediverse software systems and several kinds of relation); this package is the Lemmy federation-with-blocks series only.
The Lemmy intra-instance graph of the Fediverse as crawled by the Fedivertex project: 24 snapshots taken between 3 February 2025 and 14 July 2025. In the upstream documentation this is an intra-instance graph: instance X is linked to instance Y if a user of X has published a message on instance Y; the graph is directed and very sparse. The package holds 24,692 links, each with the snapshot it was seen in, and 9,878 instance records (557 distinct instances; one record per instance per snapshot) with the attributes upstream recorded (`version`, `users`, `posts`, `comments`, `communities`, `users_active_day`, `users_active_week`, `users_active_month`, `users_active_half_year`, `captcha_enabled`, `require_email_verification`, `description_language`). A snapshot lists between 362 and 439 instances and between 797 and 1,133 links. Edge weights are whole numbers from 1 to 86,402 (1,102 distinct values). Instance names are written with "[DOT]" in place of each dot, exactly as upstream published them. The graph is declared directed. Fedivertex covers 13 such graph series (seven Fediverse software systems and several kinds of relation); this package is the Lemmy intra-instance series only.
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.
Manufacturing emails network, a network of the Koblenz Network Collection (KONECT) (Employee–employee email). KONECT's file for it has the header `% asym positive`, i.e. a directed network with weights marked `positive`. This copy holds 82,927 links among 167 nodes, timed from 2 January 2010 to 30 September 2010.
Stack Exchange
Epinions
LiveJournal
Twitter (X)
BlogCatalog
Slashdot
Reddit
Myspace
Fediverse
Mastodon
Advogato
Wikipedia
Facebook