This post is part of a literature review on misinformation, especially in relation to technology and social media. The intention is to organize my thoughts and to provide the reader with condensed information beyond what can be found in the abstract.
For each paper I give the setting, main contributions, data used, and some strengths and limitations. These opinions are my own and are presented to give the reader some ideas. They do not represent an ironclad critical review; your judgment may vary, and so might mine on subsequent readings.
The following are from searching for the keyword “misinformation,” among 2020 WWW (and companion) papers. 23 results total. Results are ordered here according to the search results order.
Misinformation Battle Revisited: Counter Strategies from Clinics to Artificial Intelligence
Setting: high level survey of methods to counter misinformation, including education, legislation, fact-checking, rumour clinics, public inoculation, filtering, downgrading, and more.
Main Contributions: limited as it’s more a survey paper. Suggests more co-regulation (organization self-regulation + higher level government regulation) is needed to win the battle.
Strengths: general ideas and broad perspective, including history and many strategies outside ML.
Limitations: don’t expect concrete algorithms or similar. Co-regulation advocacy is not very in-depth, either in recommendations or supporting analysis.
Mitigating Misinformation in Online Social Network with Top-k Debunkers and Evolving User Opinions
Setting: counter misinformation by inducing some users (“debunkers”) to spread counter message. In contrast to previous work, here examines setting when user beliefs (in true or fake info) can vary, instead of being static once formed.
Main Contributions: algorithm to select best debunkers in above setting.
Data: “Facebook” and “Twitter” (exact sourcing unclear).
Strengths: more realistic than static opinions. Real-world experimental results detecting what opinions will be formed in a network, on a given topic, look promising.
Limitations: debunker compliance unclear in the real world. Real-world experiments are prediction only, no interventions tested. Misinformation topics must be specified by user.
Proactive Discovery of Fake News Domains from Real-Time Social Media Feeds
Setting: detect fake news websites. Input keyword(s) for topic -> extract websites on that topic from tweets using network connections -> classify fake/real using topic-agnostic classifier on the webpage structure.
Main Contributions: shows that above content-less approach is effective and provides algorithm to do it.
Data: tweets on Donald Trump impeachment, website HTML from tweet URLs, MediaBiasFactCheck.
Strengths: results look promising and system is close to real-world applicability (pipeline including dashboard, API in the works).
Limitations: requires keyword input. Experiments limited because only one topic tested.
Factoring Fact-Checks: Structured Information Extraction from Fact-Checking Articles
Setting: fact-checks lack standardization; compliance with an existing standard is low and takes substantial effort.
Main Contributions: NLP algorithm for extracting structured information from unstructured fact-check articles.
Data: fact-checks linked from DataCommons.
Not that useful at current stage of project. Skimmed only.
A Kernel of Truth: Determining Rumor Veracity on Twitter by Diffusion Pattern Alone
Setting: determine if rumor is true or false based only on how it spreads (no content, no user metadata, etc.).
Main Contributions: Shows diffusion is predictive and gives an algorithm: Weisfeiler-Lehman graph kernel to embed cascades, and then a binary classifier on the embeddings.
Data: extremely extensive twitter + fact-check data from _The spread of true and false news online. Note however the relatively easily obtainable version contains diffusion and veracity information only, no content (and is still not that easy to obtain). Anything further would require an agreement with Twitter.
Strengths: Strong results that diffusion patterns are predictive. Potentially easy extension out of diffusion-only setting (node tags with WL).
Limitations: Supervised. As authors note, paper shows importance of diffusion, rather than a solution to real-world problems, which should use more than diffusion alone.
Unveiling Coordinated Groups Behind White Helmets Disinformation
Setting: case study investigating disinformation about White Helmets (Syrian civil war). Applies two approaches to detect coordinated groups: rapid retweets and tweet similarity.
Main Contributions: analysis is fairly exploratory, but identifies some coordinated groups of 4 types: multiple accounts sharing same content, few accounts with similar tweets, factory of news websites, and accounts using automation.
Data: DARPA SocialSim
Strengths: focused case analysis of real-world misinformation.
Limitations: analysis is not very sophisticated or conclusive.
Facebook Ads Monitor: An Independent Auditing System for Political Ads on Facebook
Setting: detect which ads are political using CNN and word embeddings. Real-world application on 2018 Brazilian election.
Main Contributions: Algorithm to accomplish above. Results suggest Facebook’s political ad identification system is deficient.
Data: Facebook ads, both from Facebook Ad Library and from their own collecting.
Strengths: Important conclusion. Some successful deployment: created browser extension to collect and classify ads, presented to Brazilian senate…
Limitations: Brazilian election only. Potentially vulnerable to adversarial evasion.
VRoC: Variational Autoencoder-aided Multi-task Rumor Classifier Based on Text
Setting: rumor classification problem defined as detection, tracking, stance classification, and veracity classification.
Main Contributions: algorithm to do all the above jointly: VAE codes tweet text and the code is used in 4 BiLSTM classifiers (one per sub-problem), with a “co-train engine” that combines losses for joint training.
Data: combination of PHEME5, PHEME9, and RumourEval (tweets related to news with rumor, veracity, and stance labels).
Strengths: only “one” (combined) dataset tested, but results are strong and compare favorably to many other methods. Does all 4 tasks together while most methods only do 1.
Limitations: content only, no network. Tracker principle - classify into given groups, rather than related/not-related - could be limiting. Supervised, which is particularly problematic for tracking and veracity classification, though not fatal.
FakeFinder: Twitter Fake News Detection on Mobile
Setting: Short paper on mobile app for detecting fake news. Particularly about speed, but also brief discussion of interface and related.
Main Contributions: small model (ALBERT) running locally achieves better speed and comparable performance to large model (BERT) on cloud.
Data: tweets and comments (collection not detailed but not really important).
Strengths: indeed it does achieve better speed.
Limitations: I’m not sure about their premise. “We designed and developed a mobile app for real-time fake news detection from live Twitter streams, which is an effective approach to contain spread of fake news on Twitter.” This requires people to install the app, but will they be the people that are susceptible to fake news? Also, there is extreme overhead if multiple users checking veracity of the same thing - how would cloud compare if you can calculate results once and store them?
Social Media Dark Side Content Detection using Transfer Learning Emphasis on Hate and Conflict
Setting: identify malicious social media content in cases where data is limited, especially different languages. Broadly includes misinformation but never explicitly focused on it. Application to Amharic language.
Main Contributions: transfer learning NLP method to handle this tricky and less-studied situation.
Data: their own Facebook crawling + AmharicCorpus (public web crawling data in Amharic).
Strengths: seems to work, though a bit unclear on how the datasets are constructed. Broad applicability.
Limitations: no network or comparison with network-based or other structural methods, which could also work on different languages.
On Twitter Purge: A Retrospective Analysis of Suspended Users
Setting: descriptive analysis of purged twitter users.
Main Contributions: data collection method: collect tweets over a period of time and then identifying which ones were purged at the end of the period (because one cannot collect purged user tweets retroactively). Some insights on this data.
Strengths: data collection method enables insights on a group we don’t know much about.
Limitations: insights can be fairly surface level. Authors suggest sparsity of the data is an issue (could not collect entire tweet history of purged users, only some).
Characterizing Search-Engine Traffic to Internet Research Agency Web Properties
Setting: analysis of IRA actions using a combination of social media (Facebook, Twitter), search (Bing), and browser (Internet Explorer 11) data.
Main Contributions: use of the latter two. Highlights how a substantial chunk of IRA actions were apolitical and not misinformation, which contributed substantially to their engagement.
Search setting is different from the data we are currently working with, so directly relevant insights are limited.
Towards Detection of Subjective Bias using Contextualized Word Embeddings
Setting: detect content bias on the sentence level using BERT variants and Wikipedia edit data.
Main Contributions: Sentence level is more interesting than word level like some previous work, though they state they want to extend to document level in the future.
Data: WNC dataset (~400k Wikipedia sentences, before and after being de-biased)
Strengths: outperforms baselines.
Limitations: it’s fine for its length, but that length is very short, so don’t expect a whole lot.
The Chameleon Attack: Manipulating Content Display in Online Social Media
Setting: examines attack by editing content, e.g. making a non-controversial post, getting people to interact in some way with it (e.g. “like”), and then changing the post to something completely different. Possible, for example, with a redirect link: changing the redirect changes the link preview on the social media site.
Main Contributions: the problem itself, and ways to mitigate it.
Data: tests the attack in practice on facebook football fan groups, using it to get a fan pages for rival clubs into the groups.
Strengths: convincing analysis that this is something to be concerned about.
Limitations: more real-world data would be nice, e.g. on how prevalent this is, though of course collection is difficult.
Seeding Network Influence in Biased Networks and the Benefits of Diversity
Setting: Seeding networks, i.e. choosing nodes in the network to target to start cascades (for example for advertising, public health, etc.), with the efficiency-maximizing goal of reaching as much of the network as possible. Strategies designed purely with efficiency maximization in mind can lead to very unequal outcomes for different groups in the network. So this paper examines seeding more diversely.
Main Contribution: shows strongly strongly that seeding for diversity need not sacrifice and can even improve efficiency.
Data: ~53k computer science co-authors labeled with first-name-perceived gender.
Strengths: gives practical insight on societally important topic.
Limitations: assumes optimization takes place with aggregated metadata, e.g. we know how many connections a person has but not who they are. However, if this assumption does not hold and we know who the connections are, then rather than an optimization based purely on degree, with or without diversity, one can do a more effective optimization, e.g. greedy. And in this case their proof no longer applies, the results become uncertain, and there may be better diversity-preserving algorithms. Therefore, applicability of this result depends on the situation, i.e. the data and computation constraints.
Stop tracking me Bro! Differential Tracking of User Demographics on Hyper-Partisan Websites
Setting: examines differences in tracking using 9 constructed personas. Finds patterns using co-clustering of personas and websites. Focus on comparing right vs. left websites.
Main Contributions: the persona methodology. Right-leaning websites track more, and some similar demographic insights.
Data: 556 websites crawled with each persona, with political leaning labels.
Strengths: persona idea is novel and seems effective.
Limitations: 9 personas may be a bit few. Not sure if disabling all ad and similar blockers is representative user behavior. Right vs. left analysis is interesting but not sure how applicable it is. Analysis is descriptive only; more information on how predictive this information is would be useful.
Identifying Referential Intention with Heterogeneous Contexts
Setting: identify the intention of a content author towards a reference, between “strong accept,” “accept,” “background,” and “strong reject.”
Main Contributions: algorithm to do above by combining the referring content, the content before and after the referral, the referred content, and the network context. First embeds each separately, applies separate attention, and then combines with another global attention.
Data: academic dataset of papers on two topics with 1565 citations hand-labeled. News/social media dataset with 401 tweet references to 297 news articles.
Strengths: sophisticated model, with experimental results indicating the combination is effective.
Limitations: supervised, and labeling even by domain experts is “difficult and time-consuming.” If cross-domain (e.g. different topic than training data) performance is not good, this would be a substantial problem. Experiments are only in-domain.
Conquering Cross-source Failure for News Credibility: Learning Generalizable Representations beyond Content Embedding
Setting: identify fake news through NLP focused on syntactical content.
Main Contributions: sophisticated algorithm to do above.
Data: 3 datasets with news articles and veracity labels: Fake or Real News, Kaggle-JR, NELA-GT-2018.
Strengths: good cross-source (i.e. train on one news platform, test on another) results.
Limitations: its strength can be a weakness in that there’s a lot of information it doesn’t use, including non-syntactical content, website structure, network, etc.
OpenCrowd: A Human-AI Collaborative Approach for Finding Social Influencers via Open-Ended Answers Aggregation
Setting: crowd-source detection of influential/authoritative figures in social media, and learn how to use the crowd data.
Main Contributions: Bayesian model of the problem and worker data and an EM algorithm to optimize it.
Data: asked twitter users to name influencers, collected potential ones, hand-labeled with expert guidance, and combined with metadata from Twitter.
Strengths: substantially reduces reliance on experts and expert-labeled data (small number still needed). Good performance.
Limitations: needs crowd, and still some experts. Network aggregate information only.