A Amazon.com movie product pages dataset of 7,911,696 movie reviews spanning 1997 to 2012, collected by McAuley & Leskovec via Amazon product review pages (2013). Each record contains product ID, reviewer user ID, reviewer profile name, helpfulness votes (numerator/denominator), star rating (1-5), Unix timestamp, review summary, and full review text. One of the largest Amazon review datasets publicly available — 8 million movie reviews with full text, helpfulness votes, and star ratings spanning 15 years of Amazon movie reviews.. Suitable for sentiment analysis, review helpfulness prediction, recommendation systems, user expertise modeling, longitudinal review behavior analysis.
A Wikipedia (English edition, 2013 snapshot) dataset of 4,203,323 article nodes and 101,311,614 directed hyperlink edges, collected by Leskovec et al. via Wikipedia XML dump (SNAP, 2013). Each record contains source article ID, destination article ID (directed hyperlink). Complete English Wikipedia hyperlink graph from a 2013 snapshot — one of the largest hyperlink networks publicly available, capturing the full link structure of 4.2M Wikipedia articles.. Suitable for hyperlink network analysis, graph evolution studies, link prediction, Wikipedia structure research, large-scale directed graph algorithms.
A YouTube (friendship network) dataset of 1,134,890 nodes and 2,987,624 edges, collected by Mislove et al. via YouTube crawl (2007). Each record contains user-user friendship edges and user-defined group communities. The dataset is notable for its large-scale YouTube friendship network with ground-truth communities. Intended for research into community detection, social network analysis, and group membership prediction.
A Twitter dataset of 456,631 users, 14,855,875 follow/mention edges, collected by De Domenico et al. (2013) via Twitter API around Higgs boson announcement (July 4, 2012). Each record contains directed edges between Twitter users representing follows or mentions, capturing information propagation during a major scientific announcement. Temporal network capturing information spreading around the Higgs boson discovery announcement, enabling study of scientific rumor propagation and information cascades in real-time social media. Suitable for information propagation, rumor dynamics, temporal network analysis, scientific communication studies, social media epidemiology.
A Amazon (co-purchasing network) dataset of 334,863 nodes and 925,872 edges, collected by Yang & Leskovec from Amazon website (2012). Each record contains product-product co-purchasing edges and product-category communities. The dataset is notable for its product co-purchasing network with ground-truth communities from categories. Intended for research into community detection, recommendation systems, and product relationship analysis.
A Web (news sites and blogs, Aug 2008–Apr 2009) dataset of 71,568 phrase clusters, 310,457 unique phrase nodes, 238,889 directed root→variant edges, 71,568 phrases in text corpus, collected by Leskovec et al. via web crawl of 1.5M articles/day, August 2008 to April 2009, distributed by SNAP. Each record contains phrase cluster graph where nodes are natural language phrases extracted from news articles and blogs; directed edges from root phrase to variant phrases within each cluster; edge weight = variant phrase frequency; covers the 2008 US election, financial crisis, and other major news events. The dataset is notable for its foundational information diffusion dataset capturing how phrases mutate and propagate across mainstream media and blogs; the cluster graph encodes phrase similarity and co-occurrence structure, enabling study of linguistic variation in news cycles. Intended for research into information diffusion modeling, meme propagation analysis, phrase clustering, linguistic variation in news media, temporal network analysis, and graph-based NLP research.
A StackExchange SuperUser (technical Q&A community) dataset of 194,085 users and 1,443,339 directed interaction edges, collected by Paranjape, Benson & Leskovec via SNAP (2017). Each record contains source user ID, target user ID, Unix timestamp of interaction. Temporal directed interaction network from SuperUser covering answers, comments on answers, and comments on questions — three types of user interactions collapsed into a single edge list with Unix timestamps.. Suitable for temporal network analysis, Q&A community structure, motif detection, information flow in technical communities.
A Wikipedia (online encyclopedia) dataset of 11,377 users and 196,614 directed signed votes (143,879 support, 40,667 oppose, 12,068 neutral) with 189,541 vote rationale texts, collected by West and Leskovec from Wikipedia's public adminship election records spanning 2003–2013, distributed by SNAP. Each record contains directed signed edges from voter to candidate: +1 support, -1 oppose, 0 neutral; election outcome per candidate (0/1); natural language vote rationale text per vote. The dataset is notable for its unique combination of signed network structure with natural language rationale text per edge, enabling joint graph and text analysis; one of the few real-world signed networks with explicit textual justifications for positive and negative links. Intended for research into signed network analysis, trust and reputation modeling, text-attributed graph research, opinion mining, link sign prediction with text features, and community detection in collaborative platforms.
A StackExchange Ask Ubuntu (Linux/Ubuntu Q&A community) dataset of 137,517 users and 280,102 directed interaction edges, collected by Paranjape, Benson & Leskovec via SNAP (2017). Each record contains source user ID, target user ID, Unix timestamp of interaction. Temporal directed network of answer-to-question interactions on Ask Ubuntu — edges connect answering users to question-asking users with Unix timestamps, capturing the flow of technical knowledge.. Suitable for Q&A network analysis, temporal interaction patterns, community detection, knowledge flow in technical communities.
A Facebook (social media platform) dataset of 134,833 Facebook pages and 1,379,403 undirected mutual-like edges across 8 page categories (company, politician, tvshow, government, artist, new_sites, public_figure, athletes), collected by Rozemberczki et al. via Facebook Graph API, distributed via GEMSEC/SNAP. Each record contains undirected edges between Facebook pages representing mutual likes; pages categorized as company, politician, TV show, government, artist, news site, public figure, or athlete; node IDs namespaced by category. The dataset is notable for its multi-category Facebook page network with known ground-truth category labels per node, enabling direct evaluation of community detection and node classification algorithms across heterogeneous page types. Intended for research into community detection, node classification, graph embedding, cross-category network comparison, and social media page recommendation research.
A Epinions.com (consumer review platform) dataset of 131,828 users and 841,372 directed signed edges (717,667 trust, 123,705 distrust), collected by Richardson et al. via Epinions.com crawl, distributed by SNAP. Each record contains directed signed edges between users: +1 indicates explicit trust, -1 indicates explicit distrust, capturing both positive and negative social relationships. one of the largest publicly available signed social networks, with explicit user-declared trust and distrust labels rather than inferred sentiment — making it ideal for signed graph algorithm benchmarking. Suitable for signed network analysis, trust propagation, link sign prediction, community detection in signed graphs, and balance theory validation.
A Google+ (now defunct) dataset of 102,100 nodes and 13,548,655 directed edges across 132 ego networks with ground-truth circles, collected by McAuley and Leskovec (Stanford, 2012) from public Google+ profiles. Each record contains directed follow edges, ground-truth circle memberships (Google+ circles shared publicly), anonymized node feature vectors representing profile attributes. 133 Google+ ego networks with publicly shared ground-truth circles from users who voluntarily released their circle data, providing a uniquely clean community ground-truth from a directed social platform. Suitable for overlapping community detection, ego network analysis, directed social circle inference, link prediction.
A Slashdot.org (technology news and discussion) dataset of 82,144 users and 1,601,229 directed signed edges (1,234,967 friend, 366,262 foe) merged from 3 snapshots, collected by Leskovec et al. via Slashdot crawls over November 2008 to February 2009, distributed by SNAP. Each record contains directed signed edges between users: +1 indicates an explicit friend tag, -1 indicates an explicit foe tag; three temporal snapshots merged into one cumulative graph. one of the canonical signed social network benchmarks, notable for its explicit friend/foe labeling system and use as a primary dataset in foundational signed network research including balance theory and status theory studies. Suitable for signed network analysis, link sign prediction, structural balance theory, status theory validation, community detection in signed graphs, and trust/distrust propagation modeling.
A Twitter (ego-networks) dataset of 81,306 nodes and 1,768,149 edges, collected by McAuley & Leskovec via Twitter public crawl (2012). Each record contains directed user-user follow edges and ego-network circles. The dataset is notable for its directed ego-network dataset with ground-truth social circles. Intended for research into social network analysis, influence propagation, and community detection.
A Brightkite (location-based social network) dataset of 58,228 users, 214,078 friendship edges, and 4.49M check-ins, collected by Cho et al. via Brightkite API (2008-2010). Each record contains undirected friendship edges, check-in records (timestamp, latitude, longitude, location_id). The dataset is notable for its location-based social network with rich spatiotemporal check-in data for mobility research. Intended for research into human mobility analysis, location-based social networks, friendship prediction, and urban computing.
A Reddit dataset of 54,075 subreddits and 571,927 directed hyperlinks (510,787 positive sentiment, 61,140 negative sentiment) from post hyperlinks embedded in post titles, collected by Kumar et al. via Reddit API, January 2014 to April 2017, distributed by SNAP. Each record contains directed edges from source to target subreddit with binary sentiment label (+1/-1) and 86-dimension text property vector; post ID and timestamp per edge. captures cross-community interaction and conflict on Reddit with sentiment polarity per hyperlink, enabling study of how communities reference each other positively vs. antagonistically over time. Suitable for community conflict detection, cross-community interaction modeling, signed network analysis, temporal network analysis, and online community dynamics research.
A Reddit dataset of 35,776 subreddits and 286,561 directed hyperlinks (265,491 positive sentiment, 21,070 negative sentiment) from post hyperlinks embedded in post bodies, collected by Kumar et al. via Reddit API, January 2014 to April 2017, distributed by SNAP. Each record contains directed edges from source to target subreddit with binary sentiment label (+1/-1) and 86-dimension text property vector; post ID and timestamp per edge. captures cross-community interaction and conflict on Reddit with sentiment polarity per hyperlink, enabling study of how communities reference each other positively vs. antagonistically over time. Suitable for community conflict detection, cross-community interaction modeling, signed network analysis, temporal network analysis, and online community dynamics research.
A Twitch.tv (live streaming platform) dataset of 34,118 users and 214,556 undirected friendship edges across 6 language communities (DE, ENGB, ES, FR, PTBR, RU) merged into one cumulative graph, collected by Rozemberczki et al. via Twitch API, distributed via MUSAE/SNAP. Each record contains undirected friendship edges between Twitch streamers within each language community; node IDs namespaced by language to avoid collision across subgraphs. The dataset is notable for its multilingual streaming platform social network covering 6 distinct language communities, enabling cross-community structural comparison and multilingual graph embedding research. Intended for research into community detection, graph embedding, node classification, cross-platform social network analysis, and multilingual network comparison.
A arXiv High Energy Physics — Theory (1991–2003) dataset of 27,770 nodes and 352,807 edges, collected by J. Leskovec et al. via arXiv. Each record contains directed: True, weighted: False. Citation network of HEP-TH papers on arXiv — directed, covers theoretical physics literature.. Suitable for citation network analysis, academic influence, graph mining.
A Math Overflow (ALL) dataset of 24,818 users and 239,978 temporal edges, collected by Paranjape, Benson, Leskovec via Stack Exchange Data Dump (2017). Each record contains directed temporal edges (all) between users with Unix timestamps. The dataset is notable for its ALL interaction type from a 6.4-year temporal network of academic Q&A. Intended for research into temporal network analysis, motif discovery, social interaction dynamics, and academic community analysis.
A Gnutella (August 25, 2002) dataset of 22,687 nodes and 54,705 edges, collected by Ripeanu et al. via Gnutella crawl (2002). Each record contains directed peer-to-peer connections in the Gnutella network. The dataset is notable for its snapshot from August 25, 2002, part of a temporal sequence of 9 snapshots. Intended for research into peer-to-peer network analysis, temporal graph evolution, and network resilience studies.
A Math Overflow (A2Q) dataset of 21,688 users and 90,489 temporal edges, collected by Paranjape, Benson, Leskovec via Stack Exchange Data Dump (2017). Each record contains directed temporal edges (a2q) between users with Unix timestamps. The dataset is notable for its A2Q interaction type from a 6.4-year temporal network of academic Q&A. Intended for research into temporal network analysis, motif discovery, social interaction dynamics, and academic community analysis.
A arXiv Astrophysics (collaboration network 1992–2003) dataset of 18,771 nodes and 198,050 edges, collected by J. Leskovec et al. via arXiv. Each record contains directed: False, weighted: False. Co-authorship network of astrophysics papers on arXiv — nodes are authors, edges connect co-authors.. Suitable for collaboration network analysis, community detection, academic social networks.
A Gnutella (August 06, 2002) dataset of 8,717 nodes and 31,525 edges, collected by Ripeanu et al. via Gnutella crawl (2002). Each record contains directed peer-to-peer connections in the Gnutella network. The dataset is notable for its snapshot from August 06, 2002, part of a temporal sequence of 9 snapshots. Intended for research into peer-to-peer network analysis, temporal graph evolution, and network resilience studies.
A MOOC online learning platform dataset of 7,047 students, 97 course content items, and 411,749 student actions, collected by Kumar, Zhang & Leskovec via SNAP (2019). Each record contains action ID, student user ID, course content item ID, relative timestamp, dropout label (0/1), and 4 pre-computed action feature dimensions. Bipartite temporal network of student interactions with MOOC course content — 1.0% dropout rate across 396,633 actions. Includes pre-computed embedding features per action, enabling direct use in temporal graph learning without feature engineering.. Suitable for student dropout prediction, temporal interaction networks, dynamic graph embedding, online learning behavior analysis.
A Bitcoin OTC (peer-to-peer Bitcoin trading platform) dataset of 5,881 users and 35,592 directed weighted trust ratings (32,029 positive, 3,563 negative), collected by Kumar et al. via platform crawl, distributed by SNAP. Each record contains directed weighted signed edges between users: integer rating in [-10, +10] representing trust strength, plus binarized sign (+1/-1) for compatibility with signed network algorithms; Unix timestamps per rating. continuous integer trust weights rather than binary +1/-1 signs, enabling weighted signed network analysis and providing finer-grained trust signal than Epinions or Slashdot while operating in the high-stakes context of peer-to-peer Bitcoin trading. Suitable for weighted signed network analysis, edge weight prediction, trust propagation modeling, fraud detection, reputation system design, and signed graph embedding research.
A Wikispeedia (Wikipedia-based online navigation game) dataset of 4,586 Wikipedia article nodes, 119,767 hyperlinks, and 76,193 human navigation paths, collected by West & Leskovec via the Wikispeedia online game (2008–2012). Each record contains article hyperlink graph (source → target), navigation path sequences, path duration in seconds, path length in clicks, back-click count, and optional difficulty rating (1=easy to 5=brutal). Unique dataset combining a Wikipedia hyperlink graph with 76,193 human navigation paths from an online game where players navigate from a random start article to a target article using only hyperlinks — 67.4% of paths were successfully completed.. Suitable for semantic distance estimation, human wayfinding behavior, hyperlink graph analysis, information foraging, Wikipedia structure research.
A Facebook (ego-networks) dataset of 4,039 nodes and 88,234 edges, collected by McAuley & Leskovec via Facebook app survey (2012). Each record contains user-user friendship edges and ego-network circles. classic ego-network dataset with ground-truth social circles. Suitable for social network analysis, community detection, and circle prediction.
A Bitcoin Alpha (peer-to-peer Bitcoin trading platform) dataset of 3,783 users and 24,186 directed weighted trust ratings (22,650 positive, 1,536 negative), collected by Kumar et al. via platform crawl, distributed by SNAP. Each record contains directed weighted signed edges between users: integer rating in [-10, +10] representing trust strength, plus binarized sign (+1/-1) for compatibility with signed network algorithms; Unix timestamps per rating. continuous integer trust weights rather than binary +1/-1 signs, enabling weighted signed network analysis and providing finer-grained trust signal than Epinions or Slashdot while operating in the high-stakes context of peer-to-peer Bitcoin trading. Suitable for weighted signed network analysis, edge weight prediction, trust propagation modeling, fraud detection, reputation system design, and signed graph embedding research.
A European research institution internal email system dataset of 986 employees and 332,334 directed email interactions across 42 departments, collected by Leskovec et al. via SNAP (anonymized institutional email logs). Each record contains sender employee ID, recipient employee ID, Unix timestamp, and department label per employee (42 departments). Temporal directed email communication network from a European research institution with ground-truth department labels for all 986 employees — enabling evaluation of community detection algorithms against known organizational structure.. Suitable for community detection benchmarking, organizational network analysis, temporal email network analysis, higher-order graph clustering.
Wikipedia hoax and non-hoax articles from Kumar, West & Leskovec, 'Disinformation on the Web' (WWW 2016), distributed by SNAP. The README describes 64 hoaxes, each paired with a non-hoax article created the same day with similar appearance; the parsed folders hold those 64 + 64 articles, while the cleaned-markup folder holds 76 hoaxes and 64 non-hoaxes (140 articles). The archive is SNAP's own and also contains 65 macOS '._' resource-fork files, which are not articles. Released under the CRAPL licence (CRAPL-LICENSE.txt is included).
Amazon
Wikipedia
YouTube
Twitter (X)
Stack Exchange
Facebook
Epinions
Google
Slashdot
Brightkite
Reddit
Twitch
Gnutella