Source-linked AI summary
On the Origins of Memes by Means of Fringe Web Communities
Savvas Zannettou, Tristan Caulfield, Jeremy Blackburn, Emiliano De Cristofaro, Michael Sirivianos, Gianluca Stringhini, Guillermo Suarez-Tangil
TL;DR
The paper addresses the need to measure how memes propagate, evolve, and influence opinion across interconnected Web communities. It analyzes 160M images with perceptual hashing, clustering, Know Your Meme annotations, and Hawkes processes across Twitter, Reddit, Gab, and /pol/. The results distinguish /pol/’s large overall influence from The Donald’s higher efficiency in pushing memes across communities.
Problem
The study addresses limited evidence about meme propagation and influence across interconnected Web communities, particularly fringe communities.
Method
The authors cluster fringe-community images using perceptual hashing and a custom distance metric, annotate clusters with Know Your Meme, map other community images to them, and model influence with Hawkes processes.
Results
/pol/ has the largest overall influence for racist and political memes, while The Donald is more efficient at pushing memes to fringe and mainstream communities.
Takeaways & Limitations
The measurements provide a building block for studying meme ecosystems and developing systems to identify or protect against harmful ideological content.
Takeaways & Limitations
The study focuses on static images and selected communities, with future work extending measurements to video and additional platforms.
Abstract
from arXiv · showhide
Internet memes are increasingly used to sway and manipulate public opinion. This prompts the need to study their propagation, evolution, and influence across the Web. In this paper, we detect and measure the propagation of memes across multiple Web communities, using a processing pipeline based on perceptual hashing and clustering techniques, and a dataset of 160M images from 2.6B posts gathered from Twitter, Reddit, 4chan's Politically Incorrect board (/pol/), and Gab, over the course of 13 months. We group the images posted on fringe Web communities (/pol/, Gab, and The_Donald subreddit) into clusters, annotate them using meme metadata obtained from Know Your Meme, and also map images from mainstream communities (Twitter and Reddit) to the clusters. Our analysis provides an assessment of the popularity and diversity of memes in the context of each community, showing, e.g., that racist memes are extremely common in fringe Web communities. We also find a substantial number of politics-related memes on both mainstream and fringe Web communities, supporting media reports that memes might be used to enhance or harm politicians. Finally, we use Hawkes processes to model the interplay between Web communities and quantify their reciprocal influence, finding that /pol/ substantially influences the meme ecosystem with the number of memes it produces, while \td has a higher success rate in pushing them to other communities.
1 Introduction
The paper addresses limited measurement of meme origins, propagation, evolution, and cross-community influence, especially in fringe Web communities. It studies these questions across Twitter, Reddit, Gab, and /pol/ using large-scale image analysis and influence modeling.
- Motivation: Limited measurements and computational tools make meme origins and influence difficult to understand across interconnected Web communities.Prior work largely examined social networks in isolation.
- Research questions: The study asks how memes evolve and propagate, how their cross-community influence can be measured, how variants relate, and how communities can be characterized through memes.These questions focus on fringe Web communities.
- Approach: 160M images posted between July 2016 and July 2017 were analyzed across Twitter, Reddit, Gab, and 4chan’s /pol/.The pipeline uses perceptual hashing and clustering techniques to study meme dissemination.
- Findings: /pol/ substantially influences the meme ecosystem through the number of memes it posts, while The Donald most efficiently pushes memes to fringe and mainstream communities.Both communities are influential despite their modest size.
- Findings: Communities within 4chan, Reddit, and Gab use memes to share hateful and racist content, including variants of Happy Merchant and Pepe the Frog.These examples occur among the most popular meme clusters.
2 Methodology
The methodology identifies visually similar image clusters, groups them into memes, annotates them with external meme metadata, and maps images from Web communities to those groups. It then analyzes meme popularity, diversity, evolution, and cross-community influence.
- Overview: The pipeline clusters visually similar images to track meme variants and groups related clusters into higher-level memes.This representation distinguishes multiple clusters within a single meme, such as Smug Frog.
- Data sources: Know Your Meme, Twitter, Reddit, /pol/, and Gab provide annotation and dissemination data for the pipeline.The methodology is designed to support other annotation sites and Web communities.
- Image processing: Perceptual hashing produces 64-element image fingerprints whose Hamming distances support visually similar-image detection and clustering.The hashes are robust to image changes and reduce raw-image dimensionality.
- Clustering: DBSCAN groups images after pairwise Hamming-distance comparisons, with parallel GPU implementation supporting large image collections.The clustering operates on pHash values.
- Preprocessing: Screenshots are removed from annotation galleries because they represent social-network comments rather than meme variants.A deep-learning classifier performs this filtering step.
- Cluster annotation: Clusters are annotated by comparing their medoids with Know Your Meme images using Hamming distance and threshold θ=8.The threshold is intended to capture meme diversity while maintaining a low false-positive rate; ties use minimum average Hamming distance.
- Distance metric: The distance metric combines perceptual, meme, people, and culture similarities, with perceptual similarity modeled as exponential decay over medoid-image Hamming distance.The smoother τ controls how rapidly perceptual similarity decays.
- Distance metric: Partial-mode associates unannotated community images with known clusters using only perceptual features and the cluster medoids.This enables analysis of generic Web communities lacking annotations.
3 Datasets
The study combines four Web communities with Know Your Meme (KYM) annotation data to measure image-based memes. The datasets span mainstream and fringe platforms, with clustering applied to images from fringe communities before matching annotated clusters across all communities.
- Web Communities: The study covers Twitter, Reddit, Gab, and 4chan’s /pol/, combining mainstream and fringe Web communities.Twitter and Reddit are treated as mainstream, while Gab and /pol/ are associated with fringe communities.
- Web Communities: Twitter contributes 114M images from 1.4B parsed tweets collected between July 2016 and July 2017.The images yield 74M unique perceptual hashes.
- Web Communities: 4chan’s /pol/ contributes 4.3M archived images, yielding 3.6M unique perceptual hashes.Because threads and images are deleted after one week, the study uses the 4plebs public archive.
- Meme Annotation Site: KYM supplies meme annotations including origins, start years, descriptions, examples, tags, and higher-level categories.The study crawls 15.6K KYM entries and collects 707K associated images corresponding to 597K unique perceptual hashes.
- Meme Annotation Site: KYM entries contain 1 to 8K images, averaging 45 with a median of 9, while 28% of memes have unknown origins.YouTube, 4chan, and Twitter are the most common identified origins, at 21%, 12%, and 11%, respectively.
- Running the pipeline on our datasets: The pipeline clusters images from /pol/, The Donald, and Gab, maps KYM images and cluster medoids, then matches annotated clusters to Twitter, Reddit, /pol/, and Gab.This enables characterization of Twitter and Reddit even though they are not used for the initial clustering.
4 Analysis
The analysis measures meme diversity, annotation overlap, and branching across fringe communities using clustered images and KYM metadata. It finds substantial variation and overlap among meme variants, while highlighting the distance metric’s usefulness and limits for evolutionary analysis.
- Clusters: 12.6K annotated clusters contain 268K images from /pol/, The Donald, and Gab, supporting analysis of meme diversity and variant interplay.
- Clusters: 63%–69% of images are labeled DBSCAN noise, likely reflecting one-off images rather than recurring memes.Noise means fewer than 5 images have perceptual distance ≤8 from the instance.
- Clusters: 74% of /pol/, 70% of The Donald, and 58% of Gab annotated clusters match one KYM entry, but some clusters overlap many entries.The Conspiracy Keanu cluster matches 126 KYM entries, showing that related memes can share visual clusters.
- Clusters: A small number of KYM entries span many clusters, with Happy Merchant annotating 124 /pol/ clusters and illustrating meme recombination.
- Memes’ Branching Nature: 525 frog clusters belong to 23 memes and form four dominant categories: Apu Apustaja, Feels Bad Man/Sad Frog, Pepe the Frog, and Smug Frog.
- Memes’ Branching Nature: The custom distance metric supports phylogenetic analysis, but visually different medoids can still appear close, as with two Smug Frog variants.The reported branching analysis is limited to one frog-meme family.
4.2 Web Community-based Analysis
Across Twitter, Reddit, /pol/, and Gab, the study compares meme popularity, content categories, temporal patterns, scores, and Reddit sub-community participation. Fringe communities show more racist meme activity, while politics-related memes appear broadly across mainstream and fringe platforms.
- Meme Popularity: Pepe the Frog and its variants rank among the most popular memes on every platform, although fringe-community seeding may shape this result.
- Meme Popularity: Sad Frog appears in 4.9% of /pol/ posts, 1.3% of Reddit posts, 0.8% of Gab posts, and 0.5% of Twitter posts containing annotated images.
- Meme Popularity: Reddit and Twitter primarily share harmless or neutral top memes, whereas /pol/ and Gab show prominent hateful or racist memes including Happy Merchant.Happy Merchant appears in 3.8% of /pol/ posts and 1.1% of Gab posts containing annotated images.
- Meme Categories: 117 racist memes account for 4.4% of all memes, while 556 politics-related memes account for 21.2%.Categories are assigned from specified KYM tags, including racism, antisemitism, politics, Trump, and Clinton.
- Temporal Analysis: Racist memes are substantially more prevalent on /pol/ and Gab; /pol/ usage is steady, whereas Gab usage is bursty and increases after January 2017.Political-meme activity peaks around the 2016 elections on Reddit and /pol/, during the second presidential debate on Twitter, and later on Gab.
- Scores: On Reddit, politics-related meme posts receive higher scores than non-politics posts, while racist meme posts receive lower scores than non-racist posts; Gab shows similar political and non-political scores.Reddit means are 224.7 versus 124.9 for politics and 94.8 versus 141.6 for racist content; Gab means are 87.3 versus 82.4 for politics.
- Sub-Communities: The Donald is Reddit’s leading meme-sharing subreddit, representing 12.5% of all meme posts, 9.3% of racist-meme posts, and 26.4% of politics-related-meme posts.AdviceAnimals also ranks among the top subreddits for racist and political memes.
4.3 Take-Aways
The takeaways characterize fringe communities as important sites for political, hateful, and recombined meme activity. The analysis also indicates that meme variants can be tracked across communities, while Reddit users respond differently to political and racist meme content.
- 4.3 Take-Aways: Fringe communities use many meme variants related to politics and world events, including Donald Trump clusters across /pol/, The Donald, and Gab.
- 4.3 Take-Aways: /pol/ and Gab share hateful and racist memes more often than mainstream communities, with steady activity on /pol/ and increasing activity on Gab.
- 4.3 Take-Aways: Pepe the Frog and related neutral-seeming memes are used alongside other memes to incite hate or influence opinion about events such as ISIS and Brexit.
- 4.3 Take-Aways: The custom distance metric supports analysis of meme overlap and interplay through cluster visualizations and dendrograms.
- 4.3 Take-Aways: Reddit users rate politics-related memes highly, while racist memes receive lower Reddit scores than non-racist memes.
5 Influence Estimation
The paper uses Hawkes processes to model meme postings as interacting community processes and estimate reciprocal influence, including likely root causes. This reveals a contrast between raw influence and per-meme efficiency: /pol/ produces substantial influence through volume, while The Donald pushes memes more efficiently.
- 5.1 Hawkes Processes: Hawkes processes model each community as a process whose meme-posting events can increase subsequent event rates within or across communities.The model fits background rates, cross-community weights, and impulse shapes for each meme cluster.
- 5.1 Hawkes Processes: Root-cause probabilities are assigned among background rates and prior-event impulses according to their magnitudes when each event occurs.This attribution can include chains of influence across multiple communities rather than only one source-to-destination hop.
- 5.2 Influence: Across all clusters, /pol/ is the strongest influence source for Reddit, The Donald, and Gab, whereas Reddit most influences Twitter.The comparison uses destination-community events attributed to source communities.
- 5.2 Influence: The Donald has more than 4 times the external influence of the closest other community, while /pol/ has only 4.03% normalized external influence.Normalization measures influence relative to the number of memes posted by the source community.
- 5.2 Influence: /pol/ has the largest overall influence for racist and political memes, but it is least efficient per meme, while The Donald is most efficient.The analysis interprets /pol/ as producing many meme variants, with relatively few spreading onward.
6 Related Work
Prior work studies meme detection, popularity, evolution, generation, case studies, and fringe communities, often using isolated platforms or text-focused data. This paper extends that literature with a multi-platform image-meme analysis and Hawkes-process modeling of cross-community influence.
- Detection and Propagation of Memes: Earlier studies tracked text-based meme diffusion, clustered Reddit submissions, or analyzed politics-related content, while other work examined image dissemination during crises.These approaches cover detection and diffusion but differ in modality, platform, and application.
- Popularity of Memes: Prior popularity research modeled hashtag virality, user exposure limits, network-community features, and image traits associated with Web popularity.The cited studies span Twitter, discontinued meme platforms, and Quickmeme.
- Popularity of Memes: Unlike prior work, this study measures image-meme popularity across /pol/, Reddit, Twitter, and Gab and reports community- and ideology-dependent popularity.The Happy Merchant meme is cited as an example of popularity on /pol/.
- Evolution of Memes: Previous evolution studies found that memes develop variants and that communities diffuse particular variants, whereas this paper uses Hawkes processes to assess cross-community posting causality.The paper positions its temporal analysis as a multi-community influence model.
- Generating Memes: Research on meme generation shows that automated systems can produce recognizable memes but struggle to produce humor, while other work studies human meme creation challenges.The generation literature addresses both automated and human production.
- Case Studies and Fringe Communities: Studies of individual memes and urban legends used mixed methods or examined sharing responses, while fringe-community research documented anonymity, hate speech, and coordinated raids.The paper connects these strands to its focus on meme ecosystems involving 4chan, Gab, and Reddit subcommunities.
7 Discussion & Conclusion
The paper presents a large-scale, multi-platform measurement of meme popularity, diversity, variants, and influence across mainstream and fringe communities. It finds that /pol/ has the greatest overall influence for racist and political memes but is least efficient per posted meme, whereas The Donald spreads memes successfully across communities.
- Discussion & Conclusion: The study processes 160M images from Twitter, Reddit, /pol/, and Gab, clustering fringe-community images and mapping all-community images to annotated meme clusters.Clusters use perceptual hashing and a custom distance metric, with annotations gathered from Know Your Meme.
- Discussion & Conclusion: Meme content differs across communities: Reddit and Twitter tend toward fun memes, while Gab and /pol/ tend toward racist or political memes.The conclusion emphasizes complex relationships among memes and their variants.
- Discussion & Conclusion: /pol/ has the largest overall influence for racist and political memes but the lowest influence relative to its posting volume, while The Donald successfully pushes memes to fringe and mainstream communities.This separates raw influence from dissemination efficiency.
- Discussion & Conclusion: The authors describe the work as a first multi-platform measurement focused on fringe and potentially dangerous communities.They frame it as a building block for cultural anthropology and systems that identify harmful ideologies.
- Performance: Associating Twitter images with 12K annotated cluster medoids took about 12 days on two NVIDIA Titan Xp GPUs, or 14ms per image.The reported timing concerns the largest dataset comparison step.
- Future Work: Future work includes video memes, OCR for associated text, crowdsourced annotation improvements, and studying meme origins, variants, and effects on public opinion.These directions extend the current image-focused analysis and annotation approach.
A Clustering Parameter Selection
The clustering pipeline selects a perceptual-distance threshold by varying distances and balancing false-positive performance against the amount of noise. The authors choose a threshold of 8.
- A Clustering Parameter Selection: DBSCAN clustering uses a threshold of 8, selected by comparing clustering behavior across varying perceptual distances.The evaluation considers cluster counts, noise percentages, and false-positive performance.
- A Clustering Parameter Selection: Distances 2–4 produce substantially more noise, while distance 10 produces the least noise.Distances between 6 and 8 provide the preferred trade-off described by the authors.
- A Clustering Parameter Selection: Table 8 reports the number of clusters and percentage of noise for each tested clustering distance.The table summarizes the parameter-selection comparison.
- A Clustering Parameter Selection: Figure 17 displays the fraction of false positives in clusters across varying clustering distances.This figure supplies the false-positive side of the threshold trade-off.
B KYM and Clustering Annotation Evaluation
The paper evaluates Know Your Meme as a source for cluster labels and assesses annotator agreement and labeling accuracy. The evaluation finds substantial agreement and mostly legitimate annotations, while noting subjectivity and occasional gallery contamination.
- KYM quality: KYM is treated as the closest available Web authority on memes, although its crowdsourced content raises legitimacy concerns.The paper notes that KYM is curated but may contain incomplete pages or random gallery images.
- KYM quality: 200 sampled clusters mapped to 162 unique KYM pages, of which 3 pages (1.85%) were judged bad.The bad entries mainly lacked completeness or contained relatively many random gallery images.
- Annotation procedure: Three annotators assessed each cluster using its KYM page, medoid, and constituent images under explicit labeling guidelines.Exact gallery matches were considered correct, while unseen images were accepted when the KYM label was appropriate.
- Evaluation results: κ=0.67 indicates substantial inter-annotator agreement, and 89% of sampled clusters received a legitimate annotation to a specific KYM entry.Majority agreement was used to assess annotation accuracy.
C Screenshot Classifier
The paper develops a convolutional-neural-network classifier to identify screenshots from several social networks in the image dataset. Trained on a curated collection and evaluated out of sample, the classifier achieves strong reported performance.
- Dataset: The classifier is trained on 28.8K images depicting posts from Twitter, 4chan, Reddit, Facebook, and Instagram.The dataset combines public image sources, manually filtered images, and 10K random /pol/ images.
- Classifier architecture: The model uses two convolutional neural networks with max-pooling, a 512-unit dense layer, and a final two-unit probability output.This architecture is designed to detect screenshots containing images from the included social networks.
- Evaluation: The model is trained on 80% of the images and evaluated on the remaining 20% as an out-of-sample dataset.The evaluation uses an ROC curve together with accuracy, precision, recall, and F1-score.
- Evaluation: AUC is 0.96, with accuracy 91.3%, precision 94.3%, recall 93.5%, and F1-score 93.9%.These results are reported for the screenshot-classification model.
D Clusters examples
The clustering pipeline produces image groups corresponding to known memes and reveals substantial visual overlap among meme variants. Examples from /pol/ illustrate recurring combinations of recognizable figures and symbols.
- Cluster examples: The pipeline detects and groups images belonging to the Dubs Guy/Check Em, Nut Button, and Goofy’s Time memes.The examples shown are drawn from /pol/ clusters.
- Cluster examples: Variations of Donald Trump, Adolf Hitler, The Happy Merchant, and Pepe the Frog appear across all example clusters.The recurring elements emphasize overlap among memes rather than isolation between meme categories.
E Interesting Images
The paper presents image examples showing memes that connect political or ideological themes and can target public figures. It also documents /pol/ users’ creation and modification of memes about the authors and the paper, alongside press-created examples.
- Cross-theme connections: Images connect the Smug Frog meme with ISIS and Brexit memes, illustrating links across distinct thematic clusters.Figures 23 and 24 provide the corresponding examples.
- Political messaging: A /pol/ image depicts Hillary Clinton as Medusa and Donald Trump as Perseus to attack Clinton’s image while boosting Trump’s.The example is explicitly framed as enhancing or penalizing the public image of specific politicians.
- Memes about the authors: /pol/ users also created memes about the authors and combined a paper figure with /pol/ branding, Smug Frog, comfy pepe, and its new variant.These examples are presented as additional evidence of the community’s meme-creation ability.
- Memes about the authors: After the paper’s initial release, /pol/ users modified a “comfy pepe” image into the “very comfy Savvas Zannettou” variant.The new meme used the first author’s face and drew on a variant from the authors’ previous work.
- Press response: Journalists created memes about the paper, including an Expanding Brain meme and a Distracted Boyfriend meme.The paper describes these press examples as reflecting its quantified analysis of Internet memes.