--- license: cc-by-4.0 --- # Dataset Card for *`domain-pool`* `domain-pool` is a fine grained and cross-domain aggregate labelled set of web domains. Its default form has majority categories downsampled to present a more balanced set of 148,830 domains, while the full imbalanced set has 5,671,355 datapoints. These web domains are mainly labelled across three axes: - reliability labels form the principal category (spans all datapoints). Reliability scores can be numerical, then normalized to [0.0,1.0] where higher is better; or categorical. Categories are listed below. - factuality labels are sparser, spanning around 12k domains. They are categorical: a domain can have low, medium or high factuality, as assessed by sources like fact-checking organisations. - bias labels can, like the reliability ones, either be a continuous score on [0.0,1.0], or a category (can either be quantitative or across left/right axis). All domains also have the original data source indicated, and the dataset's domain scope (e.g., misinformation, or malware; domain-pool spans Y domains). A large part of these data sources are open-sources academic datasets, and the labels for Y domains were also collected manually from online sources (governmental, journalistic or academic) that gathered domain lists in non-machine readable format. The full composition is provided below for both dataset versions, the downsampled one and the full variant: - `domain-pool`: 5,671,880 domains, labelled across the three axes, all with at least one categorical label that can pertain to its reliability (e.g., 'fake news' or 'adult content'). Categories are listed below. - `domain-pool-downsampled`: 149,086 domains, where the dominating categories (ones with more than 20,000 datapoints) are downsampled to 15,000 or less. Due to overlaps between datasets, this brings some of the dominating categories to counts between 10 and 15,000; the processing includes an iterative optimizer that tries to minimize such loss. Downsampling here is beneficial because the entire domain pool is predominantly composed of a few large categories (e.g., adult content accounts for more than 4 million domains). ## Dataset Overview Here we present an overview of `domain-pool-downsampled`, a curated and normalized fine-grained dataset, where we downsample the top few data domains that otherwise overwhelm the dataset's composition. ### Label composition #### Reliability Reliability as a broad category encompasses three types of labels; two quantitative, and one qualitative: 1. **Continuous score ( n = 11,980 ):** these are academic-sourced float on [0.0,1.0] that explicitly relates to the domain's reliability as assessed by expert fact-checkers. 2. **3-class ( n = 12,053 ):** same type of source and meaning, these span three levels: [low, medium, high]. 3. **Categorical ( n = 149,086 ):** these are broader categories that describe the nature of the website. Most are directly related to the website's reliability (e.g. 'malware'), while some are more neutral (e.g. 'sports'). More precisely, ##### Reliability (continuous) - Count: 11,980 - mean: 0.59, Distribution: ![Reliability Distribution](https://cdn-uploads.huggingface.co/production/uploads/681e3663829118a837bbaeb3/vVNqqkuyPlqud-qkdXyUC.png) | Range | [0.0, 0.1) | [0.1, 0.2) | [0.2, 0.3) | [0.3, 0.4) | [0.4, 0.5) | [0.5, 0.6) | [0.6, 0.7) | [0.7, 0.8) | [0.8, 0.9) | [0.9, 1.0] | |--------------|------------|------------|------------|------------|------------|------------|------------|------------|------------|------------| | Domains | 49 | 161 | 809 | 1252 | 1751 | 1133 | 2415 | 2969 | 1356 | 84 | ##### Reliability (3-class) ![Value Distribution](https://cdn-uploads.huggingface.co/production/uploads/681e3663829118a837bbaeb3/nOI3VmV_xedousOtg2F9c.png) | Value | low | high | medium | |---------|------|------|--------| | Domains | 6440 | 5426 | 309 | ##### Reliability (categorical) ![Domain Distribution](https://cdn-uploads.huggingface.co/production/uploads/681e3663829118a837bbaeb3/ha-Wn2jgyfBBNG1BeojSH.png) These are the 5 largest categories; the full list can be found at the bottom of the dataset card. Within each category, the labels usually contain a flag related to reliability (e.g., within 'political', a domain can for example be labelled 'fake news' or 'fact-checker'; within 'phishing', either 'pishing' or 'not'.) | Domain / area | Count | |---------------------------|--------------| | political | 17,180 | | adult | 15,000 | | phishing | 15,000 | | gambling | 14,993 | | shopping | 14,939 | | cryptojacking, games, jobsearch, malware... | 90k+ | ##### Factuality ![Value Distribution](https://cdn-uploads.huggingface.co/production/uploads/681e3663829118a837bbaeb3/NI7Qgl_QkXaGTdqz2-vot.png) | Factuality | very low | low | medium | high | very high | |-------------|--------|------|------|----------|-----------| | Count | 230 | 2089 | 5889 | 3962 | 103 | ##### Bias (continuous) - Count: 11,477 - Mean: 0.65 ![Bias Distribution](https://cdn-uploads.huggingface.co/production/uploads/681e3663829118a837bbaeb3/FHR_Ri92DuFhFauVbuLbx.png) | Range | [0.2, 0.3) | [0.3, 0.4) | [0.4, 0.5) | [0.5, 0.6) | [0.6, 0.7) | [0.7, 0.8) | [0.8, 0.9) | [0.9, 1.0] | |--------------|------------|------------|------------|------------|------------|------------|------------|------------| | Domains | 4 | 504 | 2327 | 1592 | 2384 | 2549 | 1867 | 250 | ##### Bias (categorical) | Bias Category | Far-Left | Left | Left-Center | Least Biased | Right-Center | Right | Far-Right | Pro-Science | Pseudoscience | Conspiracy | |---------------|----------|------|-------------|--------------|--------------|-------|-----------|-------------|---------------|------------| | Domains | 23 | 305 | 757 | 966 | 969 | 483 | 270 | 118 | 256 | 202 | ### Data sources Some of the primary contributors to the dataset are: - [UT1](http://dsi.ut-capitole.fr/blacklists/index_en.php) by the University of Toulouse Capitole (88.6%), - [DQR](https://academic.oup.com/pnasnexus/article/2/9/pgad286/7258994?login=false) by Lin et al. (7.6%), - Wikipedia (3.6%), - [Lasser et al.]()'s data (3.1%). The full list: ``` ── SAMPLED POOL: 148,830 domains ── ut1 131,795 ( 88.6%) DQR 11,380 ( 7.6%) wikipedia 5,421 ( 3.6%) lasser 4,682 ( 3.1%) mbfc_raw 4,365 ( 2.9%) manual 2,128 ( 1.4%) iffy 2,001 ( 1.3%) mbfc_questionnable 1,881 ( 1.3%) checkthat 1,053 ( 0.7%) legit-phish 756 ( 0.5%) url-phish 564 ( 0.4%) zoznam 337 ( 0.2%) politifact 325 ( 0.2%) dicts 266 ( 0.2%) phish-dataset 170 ( 0.1%) sd22_approved_software 157 ( 0.1%) legal_all 154 ( 0.1%) urlhaus 150 ( 0.1%) health_all 148 ( 0.1%) ngoreport 135 ( 0.1%) paperwall 123 ( 0.1%) nelez 51 ( 0.0%) tools 43 ( 0.0%) hasbara 19 ( 0.0%) edmo_hubs 16 ( 0.0%) ``` ### Geographical Distribution Political-scoped data sources largely have country attribution. In most cases, it's the country or region that the misinformation / propaganda targets. In the case of coordinated disinformation campaigns, the perpetrators may also be attributed: | Country | Target | Perp | |----------------------------|--------|------| | USA | 3640 | 1 | | Czech Republic | 360 | 0 | | India | 348 | 52 | | Europe | 317 | 3 | | China | 241 | 206 | | Global | 221 | 0 | | United Kingdom | 190 | 0 | | Canada | 183 | 0 | | North Macedonia | 178 | 0 | | Myanmar | 97 | 0 | | Iran | 94 | 18 | | Ghana | 82 | 0 | | Ukraine | 58 | 0 | | Australia | 46 | 0 | | Georgia | 45 | 0 | | France | 38 | 0 | | Hong Kong | 36 | 34 | | Israel | 34 | 143 | | Russia | 34 | 551 | | Germany | 28 | 0 | | Africa | 27 | 0 | | Japan | 26 | 0 | | Italy | 24 | 0 | | South Korea | 24 | 0 | | Turkey | 21 | 21 | | South Africa | 20 | 0 | | Cambodia | 17 | 0 | | Taiwan | 17 | 0 | | Central African Republic | 16 | 0 | | Netherlands | 16 | 0 | | Pakistan | 14 | 0 | | Spain | 14 | 0 | | Sweden | 14 | 0 | | United Arab Emirates | 14 | 0 | | Switzerland | 13 | 0 | | Brazil | 12 | 0 | | Ireland | 12 | 0 | | Egypt | 10 | 0 | | Mexico | 10 | 0 | | Romania | 10 | 0 | | Kosovo | 9 | 0 | | Philippines | 9 | 0 | | Tunisia | 9 | 0 | | Argentina | 8 | 0 | | Austria | 8 | 0 | | Belgium | 8 | 0 | | Nigeria | 8 | 0 | | Poland | 7 | 0 | | Bangladesh | 6 | 0 | | Ecuador | 6 | 0 | | Greece | 6 | 0 | | South Asia | 6 | 0 | | Cyprus | 5 | 0 | | Denmark | 5 | 0 | | Finland | 5 | 0 | | Indonesia | 5 | 0 | | Malaysia | 5 | 0 | | Venezuela | 5 | 0 | | Bulgaria | 4 | 0 | | Kenya | 4 | 0 | | Norway | 4 | 0 | | Oceania | 4 | 0 | | Saudi Arabia | 4 | 0 | | Thailand | 4 | 2 | | Algeria | 3 | 0 | | New Zealand | 3 | 0 | | Serbia | 3 | 0 | | Singapore | 3 | 0 | | Tanzania | 3 | 0 | | Albania | 2 | 0 | | Armenia | 2 | 0 | | Chile | 2 | 0 | | Colombia | 2 | 0 | | Croatia | 2 | 0 | | Iceland | 2 | 0 | | Iraq | 2 | 0 | | Jordan | 2 | 0 | | Lebanon | 2 | 2 | | Lithuania | 2 | 0 | | North Korea | 2 | 0 | | Portugal | 2 | 0 | | Slovenia | 2 | 0 | | Sri Lanka | 2 | 0 | | Andorra | 1 | 0 | | Belarus | 1 | 0 | | Beligium | 1 | 0 | | Bosnia and Herzegovina | 1 | 0 | | Cameroon | 1 | 0 | | Costa Rica | 1 | 0 | | Cuba | 1 | 0 | | Estonia | 1 | 0 | | Guam | 1 | 0 | | Guinea | 1 | 0 | | Hungary | 1 | 0 | | Latvia | 1 | 0 | | Luxembourg | 1 | 0 | | Morocco | 1 | 0 | | Puerto Rico | 1 | 0 | | Qatar | 1 | 0 | | Syria | 1 | 0 | | Uruguay | 1 | 0 | | Zimbabwe | 1 | 0 | | Benin | 0 | 1 | ## `domain-pool` (full) The full dataset, with no downsampling, has a majority of datapoints sourced from ut1 (96.1%) and phishing datasets (1.8%, 0.8% and 0.7% for `url-phish`, `phish-dataset` and `legit-phish` respectively). It is overwhelmingly composed of adult websites (4.6 mio), phishing ones (865k) and malware (713k). ### Domain Composition ![Domain Distribution2](https://cdn-uploads.huggingface.co/production/uploads/681e3663829118a837bbaeb3/4Wc91wjxEGtTpWY0K0BMr.png) | dataset_domain | domain_count | |---------------------------|-------------:| | adult | 4,592,820 | | phishing | 865,220 | | malware | 713,376 | | jobsearch | 60,732 | | shopping | 36,960 | | games | 34,560 | | gambling | 32,233 | | political | 17,025 | | cryptojacking | 16,289 | | bank | 6,638 | | dating | 6,518 | | vpn | 6,038 | | press | 4,605 | | publicite | 4,502 | | audio-video | 3,673 | | sports | 2,361 | | coordinated campaigns | 2,329 | | blog & forums | 1,707 | | bitcoin | 1,400 | | filehosting | 946 | | manga | 735 | | social_networks | 713 | | celebrity | 666 | | drogue | 615 | | radio | 566 | | stalkerware | 525 | | educational | 510 | | financial | 470 | | webmail | 412 | | agressif | 291 | | health | 215 | | chat | 209 | | lingerie | 200 | | translation | 173 | | legal | 167 | | cult | 144 | | marketingware | 79 | | ai | 77 | | child | 75 | | cleaning | 70 | | mobile-phone | 52 | | dangerous_material | 48 | | cooking | 37 | | astrology | 29 | | sexual_education | 22 | | educational_games | 9 | | religious associations | 1 | | special | 1 | ### Data Sources ``` ── FULL POOL: 5,671,355 domains ── ut1 5,448,682 ( 96.1%) url-phish 102,468 ( 1.8%) phish-dataset 43,975 ( 0.8%) legit-phish 39,164 ( 0.7%) urlhaus 29,723 ( 0.5%) DQR 11,380 ( 0.2%) wikipedia 5,423 ( 0.1%) lasser 4,682 ( 0.1%) mbfc_raw 4,365 ( 0.1%) manual 2,195 ( 0.0%) iffy 2,001 ( 0.0%) mbfc_questionnable 1,881 ( 0.0%) checkthat 1,053 ( 0.0%) zoznam 337 ( 0.0%) politifact 325 ( 0.0%) dicts 282 ( 0.0%) health_all 215 ( 0.0%) sd22_approved_software 191 ( 0.0%) legal_all 167 ( 0.0%) ngoreport 145 ( 0.0%) paperwall 123 ( 0.0%) nelez 51 ( 0.0%) tools 46 ( 0.0%) hasbara 19 ( 0.0%) edmo_hubs 16 ( 0.0%) ``` ### Full category counts: #### Downsampled: | dataset_domain | domain_count | |----------------------------|--------------| | political misinformation | 17180 | | adult | 15000 | | phishing | 15000 | | gambling | 14993 | | shopping | 14939 | | cryptojacking | 14918 | | games | 14905 | | jobsearch | 13752 | | malware | 13680 | | bank | 6316 | | dating | 6268 | | vpn | 6030 | | press | 4524 | | publicite | 4424 | | audio-video | 3475 | | sports | 2295 | | coordinated campaigns | 2248 | | blog & forums | 1654 | | bitcoin | 1280 | | filehosting | 823 | | manga | 652 | | social_networks | 651 | | drugs | 583 | | celebrity | 565 | | radio | 549 | | stalkerware | 517 | | educational | 477 | | financial | 454 | | webmail | 406 | | agressif | 278 | | chat | 193 | | translation | 176 | | legal | 171 | | lingerie | 162 | | health | 155 | | cult | 141 | | marketingware | 77 | | ai | 72 | | child | 70 | | cleaning | 67 | | mobile-phone | 47 | | dangerous_material | 41 | | cooking | 37 | | astrology | 27 | | sexual_education | 17 | | educational_games | 8 | | religious associations | 1 | | special | 1 | - **Curated by** the CrediNet organisation, which consists of a team of collaborators from the Complex Data Lab @ Mila - Quebec AI Institute, the University of Oxford, McGill University, Concordia University, UC Berkeley, University of Montreal, and AITHYRA. - **Funding:** This research was supported by the Engineering and Physical Sciences Research Council (EPSRC) and the AI Security Institute (AISI) grant: *Towards Trustworthy AI Agents for Information Veracity and the EPSRC Turing AI World-Leading Research Fellowship No. EP/X040062/1 and EPSRC AI Hub No. EP/Y028872/1*. This research was also enabled in part by compute resources provided by Mila (mila.quebec) and Compute Canada. Data sources: - [UT1](http://dsi.ut-capitole.fr/blacklists/index_en.php) by the University of Toulouse Capitole, - [DQR](https://academic.oup.com/pnasnexus/article/2/9/pgad286/7258994?login=false) by Lin et al.), - Wikipedia, - [Lasser et al.]().