ekmpa/epidomains

Dataset Card for epi-domains 0.1.0

epi-domains is a multi-class labelled set of web domains that contribute to informational exchanges and political discussions online.

It is composed of 4 classes and totals 33,527 web domains:

• Information Integrity Actors are organizations or institutions that safeguard information integrity online, e.g., fact-checker organisations.
• Questionable Source are domains whose integrity is questionable but haven't been flagged or verified as malicious or sharing misinformation.
• Misinformation Spreaders are websites that share misinformation online.
• FIMI-related Domains are websites that have been confirmed to be linked to foreign information manipulation interference operations, or coordinated / state-led campaigns.

Dataset Variants

The dataset is made available under two variants, both being csv formats: (1) a source-attributed dataset (epidomains.csv) where all datapoints, and each of their label, is sourced; and (2) a clean version (epidomains_clean.csv) with only label values. The sourced version is also made available in the split directory with a 60-20-20 train-val-test split, stratified by classes to have each split balanced.

FIMI-related Domains:

These datapoints mainly originate from primary sources (e.g. governmental or cybersecurity reports), or require being supported by 2+ secondary sources (e.g. media relaying a claim). They are usually attributed a perpetrator and target country:

Labels

Many of the domains in this dataset also have labels, usually expert-driven scores pertaining to their reliability.

Reliability

Reliability is mainly scored as a binary flag (n = 27,750):

Though some domains also have a continuous reliability score (n = 9,020):

And a smaller part of domains also have categorical reliability labels (n = 3,433 for 3-cat, n = 1,878 for 6-cat):

Factuality

Factuality is mainly scored as a categorical dimension (n = 2,123):

Or a continuous one (n = 9,020):

Conflict Resolution

When original data sources disagree on a domain's label (for binary reliability), either (1) the label is chosen through a majority vote, (2) the domain is discarded if there are not enough sources for a majority vote.

Data sources

Some of the primary contributors to the dataset are:

• Malicious And Benign Webpages (38.8%),
• DQR (26.9%),
• Meta Threat Reports (18.8%),
• And 20+ others contributing <10%.

The full list:

Source	Rows	% of Total
Malicious And Benign Webpages Train	9,826	29.3%
Domains Quality Ratings	9,020	26.9%
Meta Threat Reports	6,319	18.8%
Malicious And Benign Webpages Test	3,172	9.5%
MBFC	2,718	8.1%
Iffy Index	1,878	5.6%
Misinfodomains	1,642	4.9%
Manual	1,410	4.2%
Hosts Fake News	1,031	3.1%
Wikipedia General	894	2.7%
CheckThat!	861	2.6%
Wikipedia Campaigns	806	2.4%
Wikipedia Miscellaneous	760	2.3%
UT1	644	1.9%
Wikipedia Fake News	402	1.2%
Zoznam	334	1.00%
Politifact's Almenac	286	0.85%
Paperwall	123	0.37%
Ngo Report Uae Blacklist	95	0.28%
Ngo Report Israel Blacklist	92	0.27%
Ngo Report Saudi Blacklist	83	0.25%
Wikipedia II Actors	60	0.18%
Nelež	51	0.15%
FakeNewsNet	48	0.14%
Hasbara Tracker	18	0.0537%
EDMO Hubs	16	0.0477%
Dns Blocklists	11	0.0328%
Ngo Report Russia Blacklist	4	0.0119%

• 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.