Dissecting the Meme Magic: Understanding Indicators of Virality in Image Memes
Chen Ling, Ihab AbuHilal, Jeremy Blackburn, Emiliano De Cristofaro, Savvas Zannettou, Gianluca Stringhini
DDissecting the Meme Magic: Understanding Indicatorsof Virality in Image Memes *Chen Ling † , Ihab AbuHilal ‡ , Jeremy Blackburn ‡ , Emiliano De Cristofaro ∓ ,Savvas Zannettou (cid:5) , and Gianluca Stringhini † † Boston University, ‡ Binghamton University, ∓ University College London, (cid:5)
Max Planck Institute for [email protected], [email protected], [email protected],[email protected], [email protected], [email protected]
Abstract
Despite the increasingly important role played by imagememes, we do not yet have a solid understanding of the el-ements that might make a meme go viral on social media.In this paper, we investigate what visual elements distinguishimage memes that are highly viral on social media from thosethat do not get re-shared, across three dimensions: composi-tion, subjects, and target audience. Drawing from research inart theory, psychology, marketing, and neuroscience, we de-velop a codebook to characterize image memes, and use it toannotate a set of 100 image memes collected from 4chan’s Po-litically Incorrect Board (/pol/). On the one hand, we find thathighly viral memes are more likely to use a close-up scale,contain characters, and include positive or negative emotions.On the other hand, image memes that do not present a clearsubject the viewer can focus attention on, or that include longtext are not likely to be re-shared by users.We train machine learning models to distinguish betweenimage memes that are likely to go viral and those that areunlikely to be re-shared, obtaining an AUC of 0.866 on ourdataset. We also show that the indicators of virality identi-fied by our model can help characterize the most viral memesposted on mainstream online social networks too, as our clas-sifiers are able to predict 19 out of the 20 most popular imagememes posted on Twitter and Reddit between 2016 and 2018.Overall, our analysis sheds light on what indicators character-ize viral and non-viral visual content online, and set the basisfor developing better techniques to create or moderate contentthat is more likely to catch the viewer’s attention.
Images play an increasingly important role in the in the waypeople behave and communicate on the Web, including emo-jis [4, 55, 93], GIFs [3, 56], and memes [27, 89, 92]. Imagememes have become an integral part of Internet culture, asusers create, share, imitate, and transform them. They arealso used to further political messages and ideologies. For in-stance, the Black Lives Matter movement [33] has made ex-tensive use of memes, often as response to racism online or to * To appear at the 24th ACM Conference on Computer-Supported Coop-erative Work and Social Computing (CSCW 2021). gather broader support [45]. At the same time, disinformationactors routinely exploit memes to promote false narratives onpolitical scandals [54] and weaponize them to spread propa-ganda as well as manipulate public opinion [21, 87, 91, 92].Compared to textual memes, image memes are often more“succinct” and, owing to a higher information density, pos-sibly more effective. Despite their popularity and impacton society, however, we do not yet have a solid understand-ing of what factors contribute to an image meme going vi-ral. Previous work mostly focused on measuring how textualcontent spreads [7, 13, 75], e.g., in the context of comment-ing [34], hashtags [90], online rumors [70], and news discus-sion [46, 47]. These approaches generally look at groups ofwords that are shared online and trace how these spread.In this paper, we set out to understand the effect that the visual elements of an image meme play in that image goingviral on social media. We argue that a popular image memeitself can be considered a successful visual artwork, created,spread, and owned by users in a collaborative effort. Thus,we aim to understand if image memes that become highlyviral present the same components of successful visual art-works, which can potentially explain why they attract the at-tention of their viewers. At the same time, we want to under-stand whether images that are poorly composed fail in catch-ing the attention of viewers and are therefore unlikely to getre-shared. Moreover, we investigate if the subjects depictedin an image meme, as well as its target audience, can havean effect on its virality. Overall, drawing from research invision, aesthetics, neuroscience, communication, marketing,and psychology, we formulate three research hypotheses vis-à-vis influence in virality.
RH1: Composition.
The theory of visual arts has constructeda comprehensive framework to characterize works of art andhow viewers process them and react to them. For exam-ple, a high-contrast color portrait of a well-depicted charactercatches viewers’ attention at first sight [66]. We argue thatimage memes may attract people’s attention in the same wayas works of art, and to investigate such hypothesis we usedefinitions and principles from the arts [16]. Although aes-thetics may contribute in a minor way to the success of animage meme, these composition features might set the basisfor an image meme to go viral. Put simply, our hypothesis is1 a r X i v : . [ c s . H C ] J a n hat an image meme that is poorly composed is unlikely to bere-shared and become viral. RH2: Subject.
Our second hypothesis is that the subject de-picted in an image meme has an effect on the likelihood ofthe meme going viral. Previous research showed that the at-tention of viewers is attracted by the faces of characters [11],and that the emotions portrayed in images capture people’sattention [83]. Therefore, we hypothesize that images that donot have a clearly defined subject do not catch the attention ofviewers and are not re-shared.
RH3: Target Audience.
Finally, we argue that the target audi-ence has an effect on virality; specifically, our intuition is thatthe audience’s understanding of a meme inherently impactstheir choice to re-share it. Thus, memes that require specificknowledge to be fully understood might only engage a smallernumber of people and will unlikely become highly viral.
Methodology.
To investigate our three research hypotheses,we follow a mixed-methods approach. We start by selectinga set of viral and non-viral image memes that were posted on4chan’s Politically Incorrect Board (/pol/), using the computa-tional pipeline developed by Zannettou et al. [92]. We choose/pol/ as previous work showed it is particularly successful increating memes that later become part of the broader Web’s“culture” [29, 59]. These images serve as a basis to iden-tify visual cues that potentially describe viral and non-viralmemes, and allow us to study our research hypotheses. Ana-lyzing this set of images and reviewing research from a num-ber of fields, we develop a codebook, detailed in Section 4, tocharacterize the visual cues of an image meme, and identifynine elements that potentially contribute to the virality of ameme. In the rest of the paper, we will refer to these elementsas features .We then have six human annotators label 100 images (50viral and 50 non-viral) according to the codebook. Our fea-tures can be easily understood by humans, as we measurehigh inter-annotator agreement, and are discriminative of theviral/non-viral classes. Finally, we train a classifier to dis-tinguish between viral and non-viral image memes on the la-beled dataset.
Main Results.
Overall, our study yields the following mainfindings:1. Our data shows that composition (RH1) does play animportant factor in determining which image memes goviral; for instance, images where the subject is takingmost of the frame are more likely to belong to the viralclass. As for RH2, we show that image memes that con-tain characters as subject are indeed more likely to goviral; in particular, this is true for characters that expressa positive emotion (for example through their facial ex-pression). For RH3, we do not find a significantly higherchance for image memes that do not require backgroundknowledge to go viral.2. Our machine learning models, built on the codebook fea-tures, are able to effectively distinguish between an im-age meme that will go viral and one that will not; the best classifier, Random Forest, achieves 0.866 Area Un-der the Curve (AUC).3. Our models trained on /pol/ image memes help charac-terizing popular memes that appear on mainstream Webcommunities too. Classifiers trained on /pol/ data cor-rectly predict as viral 19 out of the 20 most popular im-age memes shared on Twitter and Reddit according toprevious work [92].4. Through a number of case studies, we show that ourfeatures can help explaining why certain image memesbecame popular while others experienced a very limiteddistribution.Our work sheds light on the visual elements that affect thechances of image memes to go viral and can pave the wayto additional research. For instance, one could use them todevelop more effective messages when running online infor-mation campaigns (e.g., in the health context). Additionally,our models can be relied upon to factor potential virality intomoderation efforts by online social networks, e.g., by pri-oritizing viral imagery which could harm a larger audience,while discarding images that are unlikely to be re-shared.
Disclaimer.
Images studied in this paper come from 4chan’sPolitically Incorrect Board (/pol/), which is a communityknown for posting hateful content. Throughout the rest of thepaper, we do not censor any content or images, therefore wewarn that some readers might find them disturbing.
Paper Organization.
The rest of the paper is organizedas follows. Next section introduces background and relatedwork, and provide our definition of image memes used in thispaper, while in Section 3, we describe our dataset. In Sec-tion 4, we discuss our three research hypotheses regardingimage’s composition, subjects, audience, which inform thedevelopment of a codebook based on 4chan. Next, we dis-cuss the annotation process in Section 5, while in Section 6we show that the features from our codebook can be used totrain machine learning to predict virality and also generalizeon Twitter and Reddit. In Section 7, we discuss several casestudies of viral and non-viral memes across platforms; finally,we discuss the implications of our results, as well as futurework, and conclude the paper in Section 8.
In this section, we discuss background information, includingproviding the operational definition of image memes that weuse in the rest of the paper. Then, we review relevant priorwork.
The term “meme” was coined by Richard Dawkins in 1976 todesignate “any non-genetic behavior, going through the popu-lation by variation, selection, and retention, competing for theattention of hosts” [19]. Two decades later, a digital versionof memes came into the sight of academics. Sociologists firstnoticed a “new genre of cut and paste” jokes in response to the2 igure 1:
Viral “Nick Young” series meets all the conditions of a meme in this paper
Virality.
Recent research showed that image memes have be-come a prominent way for social media users to communicatecomplex concepts [21, 92]. Consistent with previous memestudies [86], we use the number of times a meme is sharedon social media as an indicator of its virality; viral memesappear at a high frequency on online services and are shared,transformed, and imitated by many people.
Image memes.
In the rest of the paper, we use the followingoperational definition. To be considered as an image meme ,an image has to meet the two following conditions:1. It must be shared by more than one user on social media;2. It must present at least one variation on social media.Figure 1 provides an example of a set of images that satisfythis definition; the example shows the “Nick Young” meme,which presents three variations.
Users are not a passive audience of media [10]. The emergingphenomenon of using memes in online communication is sup-ported by the uses and gratification theory (UGTheory) [67].The UGTheory, extended from Maslow’s hierarchy of needspyramid [52], postulates that the consumers of media choosewhat content to consume to meet their needs and achieve grat-ification. In this context, memes may function as a tool toconnect users and their audience, and be used to share infor-mation and emotions [45, 73].Moreover, different users who see a meme can behave dif-ferently, by either passively consuming it without understand-ing it, or by getting actively involved in the disseminationprocess by re-sharing or even by modifying it. Research inpsychology showed that different personality traits, such asnarcissism [48], extraversion [57], and anxiety [31] affect theway in which people use social media, particularly with re-spect to how they perceive and react to online communication. Previous research has also highlighted the role of “influ-encers” in the early spreading stage within online communi-ties [86]. More specifically, high homogeneity within onlinecommunities and the reputation of influencers induces con-formity behavior [15] in users, which helps their content goviral. A matching attitude, beliefs, and behaviors towards aviral meme can lead users to share and imitate it under un-conscious conformity to the community [15]. Additionally,research showed that individuals prefer images, or both visualand verbal information in communication [72]. Furthermore,the “pictorial superiority effect” [60] suggests that betweenan image and a textual label describing it, the image is morelikely to catch the attention of the user.
Studies on viral textual content.
Previous research studiedhow textual information goes viral on social media, and inparticular which elements influence the virality of this con-tent. A variety of features were tested to distinguish textualcontent that goes viral from the one that does not, includingcomments [46], votes [34], and user-defined groups [75, 90].Based on these features, supervised learning can be used topredict whether or not textual memes will go viral.When studying indicators of the success of textual content,previous research found that the success of online content de-pends on timing, social network structure of the account post-ing it, randomness, and many other factors [14, 85]. Otherresearch found that textual content might become viral sim-ply because it appeals to its audience [7, 13]. Another lineof work argued that the reason for virality lies on whether thecontent generates emotions (surprise, interest, or even anxi-ety/anger). However, given the competitive attention captur-ing environment of social network platforms, innate appealalone may not be able to paint the whole picture of why tex-tual memes go viral [40, 69, 90]. Finally, some researchersargue that virality is just a random event [12].
Studies on image memes.
Researchers recently startedstudying image memes and how online users interact withthem. Dupuis et al. performed a survey to identify person-ality traits of users that are more likely to share image memescontaining disinformation [21]. Crovitz and Moran provideda qualitative analysis of image memes used as a vehicle ofdisinformation on social media [17]. Zannettou et al. [92]3resented a large scale quantitative measurement of imagememe dissemination on the Web, finding that small polarizedcommunities like 4chan’s /pol/ and The_Donald subreddit areparticularly effective in pushing racist and hateful content onmainstream social media like Twitter and Reddit.
Studies on viral marketing.
Previous studies on viral mar-keting found that targeting platforms, narrative advertising,and eliciting high arousal emotion promote sharing behav-ior. Also, virality can be elicited by placing advertisementson those Web communities where the target audience is mostlikely to view and share them [62]. Online advertisementsthat trigger high arousal emotions, both positive and negative,such as anger, anxiety, exhilaration, amusement, are morelikely to be shared than those that elicit low arousal emotions,such as sadness or contentment [6, 7, 61]. Prior work in mar-keting also helps explain why people share online content. Forinstance, research suggests that affiliation is the human needto belong and form relationships [5]. Online users may there-fore share different forms of content on social media to inviteconnection and interaction with others [22, 25].
Novelty:
Overall, our work is, to the best of our knowledge,the first to investigate whether the visual features of an imagememe contribute to its virality.
To investigate the three research hypotheses highlighted inSection 1, we collect examples of image memes shared on4chan’s Politically Incorrect Board (/pol/). We choose to fo-cus on this community because previous work showed thatits users are particularly successful in influencing Internetculture and in creating image memes that will later go vi-ral [29, 59, 92]. In Section 7, we will show that, although weonly use an arguably small Web community to develop ourcodebook and identify indicators of virality, these indicatorsare in fact helpful to characterize the most viral image memesthat were shared on large mainstream Web communities likeTwitter and Reddit.We use a dataset of image memes collected by Zannet-tou et al. [92]. The authors collected a set of 160M imagesfrom Twitter, Reddit, 4chan, and Gab, and developed a pro-cessing pipeline that clusters together visually similar images.Specifically, they performed clustering on all images postedon 4chan’s /pol/, The_Donald subreddit, and Gab over thecourse of 13 months between July 2016 and July 2017. Usingground truth data from Know Your Meme [42], a comprehen-sive encyclopedia of memes, they annotated each cluster bycomparing the ground truth data with the cluster’s medoid. Finally, they mapped all images posted on Twitter, Reddit,4chan, and Gab, to the annotated clusters, hence obtaining allposts from these four Web communities that include imagememes.We obtain the 38,851 clusters that contain image memesposted on 4chan’s /pol/ from the authors of [92]. In total, thesememes appeared in 1.3M posts on /pol/. We use the number of The medoid is the point in the cluster with the minimum average distancefrom all points in the cluster. posts that contain a certain image meme as an indicator of itsvirality, i.e., we rank all image memes by the number of /pol/posts that contain them, and consider memes at the top of thelist as viral and those at the bottom as non-viral. Followingthis method, we extract two datasets:1. A sample of 100 viral and 100 non-viral clusters; specif-ically, the top 100 and the bottom 100, respectively. Weuse this dataset to identify potential indicators of viralityand build our codebook (see Section 4).2. A sample of 50 viral and 50 non-viral clusters; specif-ically, at random from the top 1,000 and bottom 1,000clusters, respectively. This dataset will be labeled byhuman annotators according to the codebook (see Sec-tion 5) and used to train machine learning classifiers totest our research hypotheses (see Section 6).For each cluster, we extract the medoid as the representativeimage meme of that cluster, thus allowing us to work on singleimages rather than clusters.Note that virality is a spectrum, with some image memesshared millions of times on social media, others only sharedonce or twice, and most falling somewhere in between. Sincethis work is the first characterization of indicators in virality ofimage memes, we believe that our choice of studying the mostpopular image memes on /pol/ and comparing them to imagesthat did not get traction is appropriate to identify indicatorsthat can tell the two classes apart. However, this is not freefrom limitations, which we discuss in Section 8.2.
In this section, we present our codebook, which guides thethematic annotation process for characterizing viral and not-viral image memes.We break the development of this codebook in two mainphases. First, one of the authors manually analyzed thedataset of the 100 most viral and 100 least viral image memesdescribed above and came up with potential indicators of vi-rality for the dataset. Next, all six authors further character-ized the image memes in the dataset with the indicators iden-tified in the first step to produce initial codes for further anno-tation, using thematic coding [9].More precisely, we followed these three steps:1. We discussed these initial codes and went through mul-tiple iterations, using a portion of the data to build a fi-nal codebook. The process continued until the codebookreached stability and additional iterations would not re-fine it further.2. To investigate the common agreement on the codebookby multiple annotators, we had them rate a portion ofour dataset and discuss disagreements until a good agree-ment is reached.3. We annotated the rest of our dataset and calculated a finalagreement.4 a) I know that feel bro (b) This is fine
Figure 2:
Example of number of panels: (a) single and (b) multiple.
Drawing from research in a number of fields, we identifya number of elements ( “features” ) that are potentially char-acteristic of image meme virality and that can help us answerour three research hypotheses. In the rest of this section, wedescribe these features in detail, along with the motivationfor selecting them, grouping them in three sections based onwhich research hypotheses they help answering: composition (RH1), subjects (RH2), or audience (RH3).
Our first research hypothesis is that the composition of an im-age contributes to making an image meme go viral. In visualarts, composition refers to the organization of visual elementsin a picture, including color, form, line, shape, space, tex-ture, and value. The different approaches to good composi-tion obey the principles of the arts, which take into accountthe balance of an image, its emphasis, movement, proportion,rhythm, unity, and variety [16]. While composition might notbe the only factor determining if an image meme will go viral(e.g., a well-designed image may not succeed as a meme be-cause users will fail to understand its meaning), we hypothe-size that a poorly composed image is unlikely to become viralin the first place.Previous research found that when viewing a scene, peoplefocus on its center first, in what is known as center bias [11,65, 78], and that the way in which objects are arranged inan image affects whether viewers perceive them as salient ornot [32]. Additionally, researchers found that movement in ascene is able to capture viewers’ attention [35].Based on previous research on image composition andmanual review of the 100 viral and non-viral memes from/pol/, we decided to use four features to investigate RH1: thenumber of panels in it, its type, its scale, and the type of move-ment in it. In the following, we describe these four features indetail.
F.1 Number of panels.
By analyzing our dataset of 200memes, we found that image memes may be composed of asingle panel or of multiple ones, similar to comic book strips.We argue that multiple panels take longer time to read than asingle image, therefore this might have an impact in gainingviewers’ attention. Therefore, we use the number of panelsin a meme to understand whether this element can affect its virality. In particular, we distinguish the following two cases:• Single panel: memes that are composed of only one im-age; e.g., see Figure 2a, “I know that feel bro,” from thefamous Wojak series, which is used in expression empa-thy or agreement to one’s expression.• Multiple panels: memes that are composed of a seriesof images; e.g., see Figure 2b, “This is fine,” originallycoming from the comic series of K.C. Green’s Gunshow,comic number 648 and used to convey hopeless emotionin a despair situation.
F.2 Type of the images.
Image memes do not only come inthe form of illustrations, but also use photographs or screen-shots, as confirmed by the sample dataset that we examined tobuild our codebook (see Section 3 and Figures 3a, 3b, and 3cfor examples). We are therefore interested in understandingwhether certain types of images better capture viewers’ atten-tion. To this end, we consider three types of images:• Photo: a picture taken by a camera (e.g., Figure 3a).• Screenshot: an image of a screenshot taken from a com-puter screen, for example of part of a Web page (e.g.,Figure 3b).• Illustration: a drawing, painting, or printed work of art(e.g., Figure 3c).
F.3 Scale.
Research in human vision showed that view-ers’ gaze is biased towards the center of a scene (i.e., centerbias) [11, 65, 78]. We hypothesize that an image that is a closeup of a subject will facilitate viewer focus on the salient partof the image and therefore catch their attention, while a largescale scene in which it is hard to identify the part to focus onmight fail in attracting the viewer’s attention. To investigatehow these aspects might affect the virality of an image meme,we consider its scale, which takes into account how the mainsubject is put in relation with the layout of the remaining ele-ments of the image. Based on the definitions of shots used infilm studies [2, 79], we define three scales for images:• Close up: a shot that tightly frames a person or object,such as the subject’s face taking up the whole frame (e.g.,Figure 4a).5 a) (b) (c)
Figure 3:
Type of images: (a) photo, (b) screenshot, and (c) illustration (a) (b) (c)
Figure 4:
Scale: (a) close up, (b) medium shot, and (c) long shot. • Medium shot: a shot that shows equality between sub-jects and background, such as when the shot is “cuttingthe person in half” (e.g., Figure 4b).• Long shot: a shot where the subject is no longer identi-fiable and the focus is on the larges scene rather than onone subject (e.g., Figure 4c).
F.4 Movement.
Research in psychonomics found that whenwatching a scene, movement is able to effectively capture theviewer’s attention [35]. Artists have been following a similarintuition for centuries, and have used a variety of techniquesto provide the illusion of movement in their paintings [23]. Inthis paper, we hypothesize that the perception of movementin an image meme might contribute to catching the viewer’sattention and therefore influence its chances to go viral.To characterize the type of movement in an image, we iden-tify three types: physical movement, emotional movement,and causal movement. A meme might contain different typesof movement at the same time. In our codebook, we considerwhether images indicate movement (as identified by arts re-search [23]), otherwise we consider them as presenting “nomovement.”We classify all movement in images as physical movement,while an emotional expression on the face or in the body lan-guage is also annotated as emotional movement. We cate-gorize movement as causal when the movement sequence iscaused by one component (sender) to another (recipient). Tak-ing Figure 5b as an example, when an initially stationary ob-ject, (the button in Figure 5b) is being set into motion by an-other moving object (the hand), viewers spontaneously inter-pret the sequence as being causal [38]. Figure 5 contains examples of the types of movement thatwe consider: (a) a meme that contains physical movementonly (people walking); (b) a meme that contains physicalmovement (the movement of the hand) and causal movement(the movement of the hand causes the movement of the but-ton, i.e., the button acts as a recipient) (c) a meme that con-tains physical movement (the character turning away from thescreen), emotional movement (the character showing a fright-ened/protective facial expression and body language with theaction of moving backwards), and causal movement (the sud-den change of the screen causes the response of the character).
Our second research hypothesis is that the subjects depictedin an image meme have an effect on whether the meme goesviral or not. In visual arts, the subject of an image refers tothe person, object, scene, or event described or represented init [71]. Basically, the subject is the essence (main idea) of thework. We hypothesize that an image that does not have a clearsubject that the viewer can focus their attention on is unlikelyto go viral. To study this hypothesis, we look at the type ofsubject (e.g., whether the focus of the image is on a characteror an object) as well as at the characteristics of this subject(e.g., their facial expression).
F.5 Type of subject.
The types of subjects that can be de-picted in an image include landscapes, still life, animals, andportraits of people [71]. Research in human vision found thatviewers’ attention tends to be attracted by the faces of thecharacters in a picture [11]. Looking at the 200 memes in ourdataset, we identify four types of subjects appearing in them:characters, scenes, creatures, and objects (e.g., see Figure 6).6 a) (b) (c)
Figure 5:
Examples of movement: (a) an image that includes physical movement, (b) an image that combines physical movement and causalmovement, (c) an image that contains physical movement, emotional movement, and causal movement. (a) (b) (c) (d)
Figure 6:
Subject of memes: (a) object, (b) character, (c) scene, (d) creature.
Images that do not contain any of these types of subjects arecategorized as “other.” More precisely, we characterize sub-jects as follows:• “Object” refers to a material thing that can be seen andtouched, like a table, a bottle, a building, or even a celes-tial body (see Figure 6a).• “Character” refers to people (see Figure 6b) or anthropo-morphized creatures/objects, such as cartoon characters.• We categorize a subject as “scene” when the situationor activity depicted in an image meme is its main focus,instead of it being on the single characters or objects de-picted in it (see Figure 6c).• “Creature” refers to an animal that is not anthropomor-phized (see Figure 6d) [63].
F.6 Attributes of the subject.
For each category, we providesubcategories to further analyze the attributes of the subjectsin the meme. For images whose subject is one or more char-acters, we consider whether the image’s visual attraction lieswith the character’s facial expression (e.g., Figure 7a) or withtheir posture (e.g., Figure 7b). In the next section (F.7), wewill describe how we further refine the character’s emotion.For the other attributes, we identify five features:• Poster Figure 7c: informative large scale image includ-ing both textual and graphic elements. There are alsoposters only with either of these two elements. Postersare generally designed to be displayed at a long-distance.• Sign Figure 7d: informs or instructs the viewer throughtext, symbols, graph, or a combination of these. • Screenshot Figure 7e: is a digital image that shows thecontents of a electronic screen display.• Scene Figure 7f: a place where an event occurs.• Unprocessed photo Figure 7g: raw photo taken by a cam-era without being modified.
F.7 Character’s emotion.
Research showed that the emo-tion perceived from images causes physiological reactionsin viewers [28] and can even speed up the perception oftime [50]. Research in neuroscience showed that faces thatexhibit an emotion capture people’s attention [83]. Based onthis research, we hypothesize that the emotions portrayed bycharacters in an image meme might have an impact on cap-turing viewers’ attention and in their decision to re-share thememe. To study this, we include a character’s emotion as oneof the features in our study. We consider images that havebeen previously annotated as “facial expression” or “posture”in F6 to be further refined, and define three states of emotions:positive, negative, and neutral (See Figure 8). All the emo-tions are annotated based on the character’s facial expressionor body language. Strong indications of emotions is annotatedas positive (e.g., Figure 8a) or negative (e.g., Figure 8b). Oth-erwise, we consider the emotion of the character as neutral(see Figure 8c). Typical positive emotions reflected in imagememes are: laughing, smiling, smug, excited, while typicalnegative emotions reflected in image memes are: crying, be-ing angry/nervous, showing impatience/boredom/shyness.
F.8 Contains words.
Image memes frequently include a wordcaption to better elicit the message of the meme (see Fig-ure 9). However, lengthy text can delay people’s recognitionof a meme [80]. We therefore hypothesize that lengthy text7 a) (b) (c) (d) (e) (f)(g)
Figure 7:
Attributes of the subject: (a) Facial expression, (b) posture, (c) poster, (d) sign, (e) screenshot, (f) scene, and (g) unprocessed photo. (a) (b) (c)
Figure 8:
Emotion: (a) positive, (b) negative, and (c) neutral. (a) (b)
Figure 9:
Examples of image memes with: (a) no words, and (b)with words. potentially impairs a meme’s virality. To better study considerthe number of words in a meme as a feature.
Our third hypothesis deals with the fact that the intended au-dience of an image meme can influence whether the memegoes viral or not. We define the audience of an image memeas the set of viewers who fully understand the meme whenthey encounter it [8]. The users of different sub-communitieshave different backgrounds and sub-cultures, and the memesthat they produce might contain elements that resonate withtheir community’s culture, but might not be easily understood by users who are not familiar with it. To understand howlimiting the intended audience of an image meme might im-pact its virality, we first attempt to determine if a meme isintended for a general audience (i.e., does not require any spe-cific knowledge) or for a particular one. Informed by previousresearch showing that hateful memes often go viral on socialmedia [92], we then aim to understand if specific tones of thememe (i.e., hateful, racist, and political tones) have an effecton its virality.
F.9 Intended audience.
We distinguish image memes intotwo categories according to their intended audience: humancommon and culture specific. Human common refers to thoseimage memes that arouse the common experiences and emo-tions of any viewer. These are supposed to be understoodby all social media users regardless of their background (e.g.,Figure 10a). Marketing research suggests that visual me-dia that uses everyday scenes and familiar situations is morelikely to resonate with a general audience [74]. In the contextof virality, we hypothesize that these memes are more likelyto go viral on social media. Culture specific memes are thosethat require some background knowledge to be fully under-stood (e.g., Figure 10b); in this case, only members of theintended community can understand the full meaning of thememe, and this can affect its virality.Based on prior work suggesting that political and hate-8 a) (b)
Figure 10: (a) The meme “manning face” as an example of “hu-man common” (b) The meme “happy merchant” as an example of“cultural specific” meme ful memes are particularly likely to be re-shared, especiallyby polarized communities [92], we further categorize culturespecific memes into hateful, racist, and political. Previousresearch argued that polarized communities like 4chan areworking on “attention hacking,” which is a way to propagatetheir idea by sharing viral memes [51].
We now discuss the methodology followed to annotate ourdataset of 100 image memes (50 viral and 50 non-viral; seeSection 3) using the codebook presented above.
Human Annotators.
The process of labeling images can-not be entirely automated for two reasons. First, while com-puter vision has made tremendous progress in automaticallyrecognizing objects in images [36, 82], image memes oftenare not well-polished pictures; rather, they include a mixtureof drawings and collages. Second, most of the features thatwe identified are highly subjective (e.g., human common vscultural specific), and it would be difficult for an automatedapproach to label them correctly. Hence, we opt to rely onhuman annotators. To this end, we developed a Web interfaceto facilitate annotation, and had six annotators use it to labelthe 100 image memes. Note, however, that we do automatethe extraction of words from the image (see F.8 in Section 4),using optical character recognition (OCR) techniques.
The majority of our codebook includes non-binary choicesand has branching depending on some of the choices selected.To address this, we built a custom codebook-oriented annota-tion platform, delivered as a Web application. The structureof a codebook is encoded as a graph, which in turn is stored asJSON in a database. Annotation questions are linked to nodesin this graph, which allows us to capture the hierarchical di-mensions of the codebook, as well as provide direct referencepoints into the codebook for annotators.It is important to note that some of our features have mutu-ally exclusive labels and others do not. Since annotators canchoose multiple labels for these features, in the end, we have 35 possible labels across 9 features for each image we anno-tate. Table 1 lists the 35 labels that we used for annotation.
Six annotators used the Web application presented above, i.e.,they were shown images from both viral and non-viral clus-ters and asked to label based on the codebook presented inSection 4. Note that three annotators were male and three fe-males, and all of them were in their 20s and 30s and had agraduate degree. The annotators have extensive experiencewith memes in general and hateful content, which makes iteasier to understand certain coded messages that might appearin a meme.
Ethical and Privacy Considerations.
Our study was ap-proved by the IRB at Boston University. More specifically,since participants were only able to choose from a set ofpre-determined options and no sensitive information was col-lected in the process, the IRB granted us an exemption.Note that human faces appear in our dataset, which mightprompt privacy concerns. However, all images with humanfaces included in this paper belong to public figures (e.g.,celebrities), therefore, we argue that the privacy implicationsof including their faces are minimal. For example, NickYoung is a famous basketball player and Peyton Manning wasa star NFL Quarterback. All other figures in the paper are ei-ther drawings or are large-scale scenes where the focus is noton the individuals. Regardless, we further discuss potentialprivacy issues in Section 8.
Inter-annotator Agreement.
Next, we measure the agree-ment among annotators using Fleiss’ Kappa score [44]. Wedo so aiming to understand if our codebook features are intu-itive and humans can reliably identify them in images, as wellas to establish whether the labeling is “reliable” enough. Notethat the Fleiss’ Kappa score ranges from 0 to 1, where 0 in-dicates no agreement and 1 perfect agreement. To determineif annotators can identify the different aspects of each feature,we break them into the corresponding possible values; for in-stance, the movement feature has five labels: physical, causal,emotional, no movement, and none of above.In Table 1, we report the Fleiss’ Kappa scores for eachfeature/option. Overall, the agreement between annotators isgenerally high: 13 of the 35 labels have almost perfect agree-ment (score above 0.8), 15 substantial agreement (score be-tween 0.6 and 0.8), and 4 moderate (score between 0.4 and0.6). Only 3 labels fall below the threshold of moderate agree-ment, specifically, “Emotional movement,” “No movement”and “Hateful.” For “Emotional Movement” and “No Move-ment,” we believe that the agreement is low because, althoughour codebook has a definition for it, the perception of emotionis subjective, and different annotators can perceive the sameimage as emotional or not. Thus, “Emotional Movement” hasthe lowest agreement among all options. For the hateful op-tion, different cultural backgrounds in the annotators may re-sult in different understandings of what constitutes a hatefulmeme.
Number of Words.
As discussed in Section 4, our codebook9 eature Label Fleiss Feature Label Fleiss
F.1 Number ofpanels F.6 Attributes ofsubjectA single panel 0.958 *** Facial expression 0.709**Multiple panels 0.958 *** Stationary pose/posture 0.649 **F.2 Image type Poster 0.674 **Photo 0.926 *** Sign 0.640 **Illustration 0.947 *** Screenshot 0.933 ***Screenshot 0.951 *** Situation 0.674 **None of the above 1.000 *** Unprocessed photo 0.772 **F.3 Scale Other 0.624 **Close up 0.666 ** F.7 EmotionMedium shot 0.813 *** Positive 0.794 **Long shot 0.895 *** Negative 0.765 **F.4 Movement Neutral 0.627 **Physical Movement 0.724 ** F.9 AudienceEmotional Movement 0.125 Human Common 0.719 **Causal Movement 0.509 * Cultural Specific 0.729 **No movement 0.366 Hateful 0.333F.5 Type ofsubject Political 0.876 ***An object/objects 0.592 * Racist 0.805 ***A character/characters 0.768 ** None of above 0.837 ***A scene/scenes 0.544 *A creature/creatures 0.457 *None of above 1.000 ***
Table 1:
Inter-annotators agreement for each feature [44]. *** indicates almost perfect agreement (Fleiss score above 0.8), ** substantialagreement (Fleiss score between 0.6 and 0.8), while * indicates moderate agreement (Fleiss score between 0.4 and 0.6).
Figure 11:
Distribution of the number of words in viral and non-viral memes. takes into account the presence of words in an image meme.In particular, we hypothesize that image memes containingtoo many words might be less likely to go viral. Unlike otherfeatures, which need to be manually annotated, we can au-tomatically extract the words in an image meme using OCRtechniques; specifically, we use the Optical Character Recog-nition (OCR) API from Google vision [58].Having extracted the distribution of words across the vi-ral and non-viral memes, we set to determine a threshold ofthe number of words that is likely to impair the virality of ameme. Later on, we will encode this threshold into a featureused by a machine learning classifier. In Figure 11, we plot the cumulative distribution functions (CDFs) of the number ofwords in viral and non-viral memes in our dataset. To assesswhether the difference in the distributions of the two classesis statistically significant, we run a two sample KolmogorovSmirnoff test [53]. The test allows us to reject the null hy-pothesis that the number of words in an image meme does nothave an effect on its virality ( p < . ), suggesting that thisis indeed a good feature to characterize viral and non-viralmemes. As we can see, the presence of words is less commonin the viral class, with 50% of the viral memes in our datasetnot containing any words. In general, the viral memes in ourdataset contain less words than the non-viral ones: 94.0% ofthem contain less than 15 words, while 31.91% of the non-viral memes exceed this number. As a result, we select 15 asthe threshold of words beyond which the virality of a mememight be impaired. We now investigate whether the features from our codebookcan be used to train machine learning models to determinewhether or not an image meme will go viral. More specif-ically, we train several classifiers to identify which memeswill go viral, and show that our features can indeed identifyviral memes. To address the limitations of tree-based clas-sifiers over limited training data (see Section 6.4), we reportthe results of both the best classifier and the best non-tree-based classifier. The best classifier is Random Forest, with10 lassifier
AUC std acc precision recall f1-scoreRandom Forest 0.87 0.13 0.98 0.98 0.98 0.98Ada boost 0.85 0.11 0.91 0.92 0.91 0.91KNN 0.83 0.10 0.98 0.98 0.98 0.98SVM 0.81 0.10 0.71 0.75 0.69 0.70Logistic regression 0.81 0.10 0.77 0.77 0.77 0.77Gaussian Bayesian 0.77 0.18 0.69 0.76 0.69 0.67Decision tree 0.75 0.18 0.82 0.84 0.82 0.82Neural network 0.79 0.10 0.90 0.90 0.90 0.90
Table 2:
Performance of all classifiers.
AUC 0.866, and the best non-tree-based classifier is KNN,with AUC 0.828. We then discuss the importance of thefeatures identified by our classifiers, reasoning on which ele-ments are particularly indicative of virality/non-virality.
We select the features for our classification models as follows.We start with the features described in Section 4; we considerthose whose options are exclusive (e.g., the scale of a memecan only be either close up, medium shot, or long shot) asa single feature, while we split those that allow multiple op-tions into multiple features (e.g., an image meme can presenta character and an object at the same time). This yields 30features: number of images; type of images (photo, illustra-tion, screenshot, none of above); scale of images; movement(physical movement, emotional movement, causal movement,there is no movement); content (an object/objects, a char-acter/characters, a scene/scenes, a creature/creatures, noneof above); type (facial expression, stationary pose/posture,poster, sign, screenshot, scene/situation, unprocessed photo,other); emotion; words; audience; cultural specific (hateful,political, racist, none of above).We then select eight classifiers to train using our features:Random Forest [49], Support Vector Machines (SVM) [76],Logistic regression [30], K-Nearest Neighbors (KNN) [84],Ada boost [24], Gaussian Bayesian [88], Decision tree [68],and Neural network [26]. For each classifier, we take the an-notated set of 100 image memes and perform a 10-fold crossvalidation, i.e., randomly dividing the dataset into ten sets andusing nine for training and one for testing. We repeat this pro-cess 10 times, and calculate the Area Under the ROC Curve(AUC) as well as its standard deviation. The results are re-ported in Table 2. We observe that Random Forest achievesthe best performance, with an AUC of 0.866. Among non-tree based classifiers, KNN has the best classification perfor-mance, achieving an AUC of 0.828.
We now set to understand which features particularly con-tribute to the classification decision. To identify indicatorsof virality (or non-virality), we perform a feature analysis ofthe best performing classifier (Random Forest). In the follow-ing, we discuss the top five features learned by this model,and whether images that present them are more likely to goviral or to not be shared. 1.
Facial expression:
The Random Forest classifier picksup the facial expression of a character as the most impor-tant feature for classification. 42 out of 50 viral memes(84%) are labeled as “facial expression” while 19 out of50 non-viral ones (38%) possess this feature.2.
Character emotion:
Image memes showing a positiveemotion are more likely to go viral (with 39% of the vi-ral memes in our dataset presenting this feature vs. 15%of the non-viral one). Similarly, a negative emotion con-tributes to virality (with 27% of the viral memes present-ing this feature compared to 17% of the non-viral ones).3.
Character posture:
Our model puts the posture of acharacter as the third most important feature. 46% vi-ral memes presenting this feature compared to 30% non-viral memes.4.
Image scale:
While a medium shot scale does not seemto affect virality in either direction (with 61% and 58%of the viral and non-viral memes in our dataset present-ing this scale), we find that image memes that present aclose up scale are more likely to go viral (34% of the vi-ral memes in our dataset present this feature, while only14% of the non-viral ones), while images that use a largeshot scale are less likely to be viral (27% of the non-viralmemes present this feature, while only 4.6% of the viralones).5.
Character as subject:
The presence of a character is themost strong indicators among other subjects in predict-ing virality. 93% of the image memes in the viral classare labeled as “character,” compared to 67% in the non-viral class.
We now re-evaluate the three research hypotheses after ana-lyzing the results of our evaluation. Our work started by set-ting out three research hypotheses on whether the composi-tion (RH1), the subjects (RH2), and the audience (RH3) of animage meme have an effect on its chance of going viral. Wehave then developed a codebook and a number of features tohelp us investigate these three hypotheses, and analyze themby training machine learning models and use them to identifythe features that are important in distinguishing between viraland non-viral memes. Next, we summarize the main findingsfrom our analysis and discuss whether they confirm our re-search hypotheses.
RH1: Composition.
Our classification model showed thatthe scale of an image is highly discriminative, with imagesthat use a close up being more likely to go viral, while thosethat use a long shot being more likely to not get shared. Thissuggests that the composition of an image does have an effecton whether an image meme will go viral or not, confirmingRH1.
RH2: Subjects.
Our classification model shows that imagememes that contain characters are more likely to go viral andthat those that contain objects are less likely to go viral. It11lso finds that image memes that contain a facial expressions(both positive and negative) are more likely to go viral, as wellas those images where the character has a particular posture.This confirms that the subjects of an image do play a big rolein whether the image meme will go viral, confirming RH2.
RH3: Audience.
We hypothesized that the target audience ofan image meme influences the chances of it to go viral. How-ever, we do not find a confirmation of this in our classificationmodel, since none of the audience-related features are foundto be important. Therefore, we are unable to confirm RH3.
Our results so far are based on a dataset of viral and non-viralimage memes shared on 4chan’s Politically Incorrect board(/pol/). Next, we want to understand if the features identifiedby our model are helpful in characterizing viral image memeson mainstream platforms. To do so, we collect an additionaldataset of the top 10 memes shared on Twitter and Reddit ac-cording to previous work [92], for a total of 20 images (seeTable 3). We repeat our annotation process on these images,and run our models trained on 4chan data on the resultingdataset, to investigate whether the indicators of virality iden-tified while training our model on /pol/ data are also presenton the most popular image memes on Twitter and Reddit.Our models are able to correctly classify 19 out of the 20most popular image memes on Twitter and Reddit as viral,indicating that our indicators of virality indeed generalize toother platforms. In Section 7, we provide a more detailed dis-cussion of some of these popular memes and of the predictorsthat allow our models to identify them as viral. We also fur-ther discuss the only meme that is not correctly predicted byour model as viral, reasoning about the reasons for this mis-classification.Note that not all classifiers perform well when trained on4chan data and tested on Reddit and Twitter one. In fact, ourbest performing classifier for this experiment is KNN. Tree-based classifiers, in particular, provide worse results than onthe 10-fold cross validation. This is a known limitation ofthese classifiers when dealing with limited training data. In anutshell, random forests can suffer from high variance acrossindividual trees, and even aggregating across these individualtrees, the strength of random forests in general, can lead toinconsistency [77].
In this section, we discuss a number of case studies to il-lustrate how our models can help evaluating viral and non-viral image memes. We first look at a viral meme (i.e.,the Smug Frog), highlighting the important features that ourmodel picks up and that can explain its virality. We then focuson a number of non-viral memes, looking at what features inour codebook might have had an effect.Finally, we analyze the two most viral memes on Twitterand Reddit from Section 6.4, showing that the same indicatorsof virality that we identified for image memes on /pol/ apply
Figure 12:
Smug Frog. to these images as well. We also analyze the only image memeamong the top Twitter and Reddit ones that our model fails topredict as viral, reasoning about why that is the case. “Pepe the Frog,” originally a comic book character createdby Matt Furie, has spawned numerous derivatives and has be-come one of the most popular meme characters online. Inparticular, polarized communities have appropriated this char-acter, to the point that it was declared as a hate symbol by theAnti-defamation League [1]. The Smug Frog (see Figure 12)is one of the many incarnations of Pepe, which originated on4chan in 2011, and has since been viral, with many variationsposted across social media. Previous work which quantita-tively measured the occurrence of memes [92] found that thismeme was prominently discussed online, appearing in 63,447threads on 4chan, 2,197 on Twitter, 392 on Gab, and 5,968 onReddit.The Smug Frog presents several characteristics that ourmodel identified as being typical of viral memes. The illus-tration uses a close up scale, tightly framing the character’sface. The character presents a positive emotion and a posture,which are all important features of virality according to ourmodel. This shows that our model is able to correctly char-acterize this image meme, identifying the elements that mighthave contributed to its virality.
Our models have identified multiple traits that help an imagememe go viral. We also found that if an image meme lacksthese characteristic traits of virality, it is unlikely to go viral.Consider for instance the four examples of memes that did notgo viral on /pol/ displayed in Figure 13.At a first glance, the images lack the important indicatorsof virality mentioned before. Figures 14a and 14b lack char-acters as subjects, and therefore cannot rely on their postureor displayed emotions to bring their message across. WhileFigure 13c and 13d do display characters, it is not immedi-ately clear to the viewer where to focus their attention, andboth memes require the viewer to read the entire text to un-derstand the point of the meme. This goes against what ourmodel has learned, that viral memes usually rely on direct vi-sual cues and few, if any, words. In fact, all four non-viral12 eddit TwitterMeme
Manning Face 12,540 (2.2%) Roll Safe 55,010 (5.9%)That’s the Joke 7,626 (1.3%) Evil Kermit 50,642 (5.4%)Feels Bad Man/ Sad Frog 7,240 (1.3%) Arthur’s Fist 37,591 (4.0%)Confession Bear 7,147 (1.3%) Nut Button 13,598 (1,5%)This is Fine 5,032 (0.9%) Spongebob Mock 11,136 (1,2%)Smug Frog 4,642 (0.8%) Reaction Images 9,387 (1.0%)Roll Safe 4,523 (0.8%) Conceited Reaction 9,106 (1.0%)Rage Guy 4,491 (0.8%) Expanding Brain 8,701 (0.9%)Make America Great Again 4,440 (0.8%) Demotivational Posters 7,781 (0.8%)*Fake CCG Cards 4,438 (0.8%) Cash Me Ousside/Howbow Dah 5,972 (0.6%)
Table 3:
Top 10 viral memes posted on Reddit and Twitter and their respective post counts between July 2016 and July 2017 (obtainedfrom [92]). Note that entries are links to corresponding Know Your Meme entries. ∗ The meme CANNOT be detected as viral in ourclassifiers, see detail explanation in Section 7.3. (a) (b) (c) (d)
Figure 13:
Non-viral memes: (a) United Airlines Passenger Removal, (b) A journalist condemn fake media, and (c) Problem solved, (d)Hang up the phone. memes in Figure 13 rely on the user reading and understand-ing the rather elaborate text in order to understand the meme.
As reported in Table 3, the most shared viral meme on Redditis Manning Face (see Figure 14a). According to [92], 12,540posts on Reddit included it between 2016 and 2017, i.e., 2.2%of all posts on Reddit. This image meme depicts NFL quar-terback Peyton Manning wearing a black hoodie, and is oftenused as a reply in Reddit threads as a bait-and-switch joke.Analyzing the meme according to our model, Manning Facepresents several features that are indicative of viral memes. Itis a close up of a character presenting a dramatic facial ex-pression.The most popular meme on Twitter is Roll Safe (see Fig-ure 15a). According to [92], 55,010 tweets included Roll Safebetween 2016 and 2017, which represent 5.9% of all imagememes posted on Twitter. Roll Safe was also among the top10 popular meme on Reddit (4,523 posts, 0.8%). This memedepicts actor Kayode Ewumi pointing to his temple, and isusually accompanied by witty text denoting smart thinking.In Figure 15b, a user captioned the meme with “If you’re al- ready late.. take your time.. you can’t be late twice.” Similarto Manning Face, Roll Safe presents many of the features thatwe learned being indicative of viral memes: it depicts a char-acter in a close up scale, presenting a positive facial expres-sion and a particular posture.Our classifiers fails to detect one viral meme from Reddit:“Fake CCG Card” (see Figure 14b), which ranks 10th on Red-dit [92]. This image meme has small text, does not really usecomposition to attract attention, does not depict emotion, etc.,and thus is a clear false negative for our model. However, a bitmore explanation about this particular meme can help reasonabout why our model failed. The “Fake CCG Card” memedepicts a modified version of a playing card from CollectibleCard Game (CCG), e.g., Magic The Gathering or Yugioh.This particular instance is actually a composition of the FakeCCG Card meme and another meme, “THE GAME,” whichyou lose by thinking about the phrase “the game.” As can beseen in Figure 14b, by viewing this “The Game trap card,”you have indeed lost “THE GAME.” “THE GAME” memeitself is quite a bit older than most memes, having originatedin the real world in the 1990s, and while this instance is defi-nitely an image meme, the “meta” aspect of “THE GAME” is13 a) Manning Face (1st) (b) Fake CCG Cards (10th)
Figure 14:
Top memes on Reddit. (a) Original Version (b) Popular Derivative
Figure 15:
Top meme on Twitter: Roll Safe. a likely explanation of why it is not captured by our model.Overall, these three case studies show that, although ourcodebook and models were developed with data from /pol/,they are helpful in characterizing memes from Twitter andReddit too.
In this work, we studied what features are indicative of imagememes going viral, across three dimensions: their composi-tion, subjects, and intended audience. We found that certainfeatures like the scale of an image, the presence of characters,of facial emotions, and of poses are particularly indicative ofvirality. Overall, we are confident that our work will encour-age further analysis of the role of images in online discussion,online abuse, and content moderation by the CSCW commu-nity.In the rest of this section, we discuss the design implica-tions of our results as well as their limitations, also promptingfuture work directions.
Our work provides evidence that aesthetic properties,grounded in art theory and neuroscience findings, have pre-dictive power in understanding whether memes will go viral.This finding has clear implications with respect to marketingand outreach. For instance, our models could be used to de-sign more effective messaging in online ad or activism cam-paigns and product promotion to increase reach.Alas, a potential worrying implication is that they couldhelp optimize the production of harmful memes, e.g., tospread disinformation [87, 91] and hateful content [39, 92].Considering that there are tools to automatically generate im-ages with text memes [20] and overall visual content [37],or even so-called cheap fakes [64], adding to the understand-ing of memes has potential negative consequences. How-ever, in this context, our models could be used by onlineservices to prioritize moderation, for example by focusingon those memes that are more likely to go viral and there- fore cause the most damage on the platform. As recent re-versals of automated-moderation systems have indicated [81]fully-automated, machine learning based techniques are stillmeaningfully inadequate compared to human moderators.Nonetheless, human moderators need tools to even hope tocombat things at Web scale; thus, we believe that integratingour findings into the toolbox that human moderators use canhelp prioritize and reason about the decisions they make.Finally, we also acknowledge that designing automatedtools that help determine whether pictures of individualsmight become viral memes can have personal consequences,including related to one’s privacy and safety. Therefore, fur-ther interdisciplinary work is needed to inform precautionarymeasures as well as meaningful informative feedback for so-cial platform operators and users alike.
Unsurprisingly, our study is not free of limitations. First ofall, virality is a spectrum. While it is easy to label imagesthat are shared tens of thousands of times as viral and thosethat are only shared once or twice as non-viral, most imagememes fall somewhat in between. In this paper, we chooseto focus on the most and least viral image memes shared on/pol/. We argue that since this is the first study of indicatorsof virality of image memes, this choice allows us to get a rea-sonable understanding of whether visual features allow us todistinguish between the two classes and answer our researchquestions. It is, however, possible that some of these indica-tors also apply to image memes that are somewhat viral, butnot the most popular. We hope that our work will encouragethe CSCW community to pursue more research in this space,attempting to answer more nuanced research questions.While our work shows that the indicators of virality learnedby our model from /pol/ image memes generalize to the topmemes posted on mainstream communities like Twitter andReddit, each online community has its own culture, and itis therefore possible that additional, community-specific indi-cators could more accurately predict image virality on thoseplatforms.14s part of future work, we plan to extend our analysis toimage memes posted on multiple platforms. In particular,it would be interesting to understand how the characteristicsand backgrounds of different communities influence the vi-rality of image memes posted on them, as well as to inves-tigate how viewers with different cultural backgrounds (e.g.,from different countries) understand the same image memes,and whether this understanding influences their decision tore-share it.An additional limitation is that posts on /pol/ are ephemeraland get periodically deleted [29]. Our definition of viralityconsiders the number of posts shared on /pol/ that include ameme, but it is not clear whether ephemerality affects howpeople share memes. We plan to investigate whether memeson platforms that are not ephemeral follow sharing patternsthat are significantly different from the ones on 4chan, as wellas study the use of other indicators of virality, such as numberof likes and retweets.Finally, online culture changes over time, thus, an openquestion is how stable our indicators are over time, andwhether our model would need to be re-trained frequently. Weargue that our features are not content-dependent, but ratherfocus on composition, types of subjects, and audience. There-fore, we do not expect the model to significantly change overtime, in a similar way in which the composition rules for “vi-ral” marketing are not changing. Nonetheless, we hope thatfuture work will study how the characteristics of viral memeschange over time, with the goal of identifying even more ac-curate indicators of virality.
Acknowledgments.
We would like to thank the anonymousreviewers for their insightful comments that helped us im-prove this paper. We also want to thank Yang Di, Yue Wang,and Yuxi Hong, who helped annotating images. This workwas supported by the National Science Foundation underGrant 1942610.
References [1] C. Anderson and M. Revers. From counter-power to counter-Pepe: The vagaries of participatory epistemology in a digitalage.
Media and Communication , 6(4), 2018.[2] D. Arijon.
Grammar of the film language . Silman-James Press,1991.[3] S. Bakhshi, D. A. Shamma, L. Kennedy, Y. Song, P. De Juan,and J. Kaye. Fast, cheap, and good: Why animated GIFs en-gage us. In
ACM CHI , 2016.[4] F. Barbieri, G. Kruszewski, F. Ronzano, and H. Saggion. Howcosmopolitan are emojis? Exploring emojis usage and meaningover different languages with distributional semantics. In
ACMMultimedia , 2016.[5] R. F. Baumeister and M. R. Leary. The need to belong: desirefor interpersonal attachments as a fundamental human motiva-tion.
Psychological Bulletin , 117(3), 1995.[6] J. Berger. Arousal increases social transmission of information.
Psychological Science , 22(7), 2011.[7] J. Berger and K. L. Milkman. What makes online content viral?
Journal of Marketing Research , 49(2), 2012. [8] L. K. Börzsei. Makes a Meme Instead.
The Selected Works ofLinda Börzsei , 2013.[9] V. Braun and V. Clarke. Using thematic analysis in psychology.
Qualitative Research in Psychology , 3(2), 2006.[10] J. Bryant and M. B. Oliver.
Media effects: Advances in theoryand research . Routledge, 2009.[11] G. T. Buswell.
How people look at pictures: a study of thepsychology and perception in art
International Workshop on Multimedia Data Mining ,2010.[14] D. Centola. The spread of behavior in an online social networkexperiment.
Science , 329(5996), 2010.[15] R. B. Cialdini and N. J. Goldstein. Social influence: Compli-ance and conformity.
Annual Review of Psychology , 55, 2004.[16] R. G. Collingwood.
The principles of art , volume 11. OxfordUniversity Press, 1958.[17] D. Crovitz and C. Moran. Analyzing Disruptive Memes inan Age of International Interference.
English Journal , 109(4),2020.[18] P. Davison. The language of internet memes.
The Social MediaReader , 2012.[19] R. Dawkins et al.
The Selfish Gene . Oxford University Press,1976.[20] A. Dubey, E. Moro, M. Cebrian, and I. Rahwan. MemeSe-quencer: Sparse Matching for Embedding Image Macros. In
WWW , 2018.[21] M. J. Dupuis and A. Williams. The Spread of Disinformationon the Web: An Examination of Memes on Social Networking.In
IEEE SmartWorld , 2019.[22] N. B. Ellison, C. Steinfield, and C. Lampe. The benefits ofFacebook “friends:” Social capital and college students’ useof online social network sites.
Journal of Computer-MediatedCommunication , 12(4), 2007.[23] C. Gottlieb. Movement in painting.
The Journal of Aestheticsand Art Criticism , 17(1), 1958.[24] T. Hastie, S. Rosset, J. Zhu, and H. Zou. Multi-class adaboost.
Statistics and its Interface , 2(3), 2009.[25] J. L. Hayes and K. W. King. The social exchange of viral ads:Referral and coreferral of ads among college students.
Journalof Interactive Advertising , 14(2), 2014.[26] R. Hecht-Nielsen. Theory of the backpropagation neural net-work. In
Neural networks for Perception . Elsevier, 1992.[27] T. Highfield and T. Leaver. Instagrammatics and digital meth-ods: Studying visual social media, from selfies and GIFs tomemes and emoji.
Communication Research and Practice ,2(1), 2016.[28] C. H. Hillman, K. S. Rosengren, and D. P. Smith. Emotion andmotivated behavior: postural adjustments to affective pictureviewing.
Biological psychology , 66(1), 2004.[29] G. E. Hine, J. Onaolapo, E. De Cristofaro, N. Kourtellis,I. Leontiadis, R. Samaras, G. Stringhini, and J. Blackburn.Kek, cucks, and god emperor trump: A measurement study of4chan’s politically incorrect forum and its effects on the web.In
ICWSM , 2017.
30] D. W. Hosmer Jr, S. Lemeshow, and R. X. Sturdivant.
AppliedLogistic Regression . John Wiley & Sons, 2013.[31] M. Indian and R. Grieve. When Facebook is easier than face-to-face: Social support derived from Facebook in socially anxiousindividuals.
Personality and Individual Differences , 59, 2014.[32] L. Itti and C. Koch. Computational modelling of visual atten-tion.
Nature Reviews Neuroscience , 2(3), 2001.[33] S. J. Jackson and B. Foucault Welles.
Infor-mation, Communication & Society , 19(3), 2016.[34] S. Jamali and H. Rangwala. Digging digg: Comment mining,popularity prediction, and social network analysis. In
WISM ,2009.[35] G. Johansson. Visual motion perception.
Scientific American ,232(6), 1975.[36] A. Karpathy and L. Fei-Fei. Deep visual-semantic alignmentsfor generating image descriptions. In
CVPR , 2015.[37] T. Karras, S. Laine, and T. Aila. A Style-Based Generator Ar-chitecture for Generative Adversarial Networks. In
Proceed-ings of the IEEE/CVF Conference on Computer Vision and Pat-tern Recognition (CVPR) , 2019.[38] D. Kerzel, H. Bekkering, A. WohlschlaÈger, and W. Prinz.Launching the effect: Representations of causal movements areinfluenced by what they lead to.
The Quarterly Journal of Ex-perimental Psychology: Section A , 53(4), 2000.[39] D. Kiela, H. Firooz, A. Mohan, V. Goswami, A. Singh, P. Ring-shia, and D. Testuggine. The Hateful Memes Challenge: De-tecting Hate Speech in Multimodal Memes. arXiv:2005.04790 ,2020.[40] M. Kitsak, L. K. Gallos, S. Havlin, F. Liljeros, L. Muchnik,H. E. Stanley, and H. A. Makse. Identification of influentialspreaders in complex networks.
Nature Physics , 6(11), 2010.[41] M. Knobel and C. Lankshear. Online memes, affinities, andcultural production.
A New Literacies Sampler , 29, 2007.[42] Know Your Meme. https://knowyourmeme.com, 2020.[43] G. Kuipers. Where was king kong when we needed him ?
TheJournal of American Culture , 28(1), 2005.[44] J. R. Landis and G. G. Koch. The measurement of observeragreement for categorical data.
Biometrics , 1977.[45] C. W. Leach and A. M. Allen. The social psychology of theBlack Lives Matter meme and movement.
Current Directionsin Psychological Science , 26(6), 2017.[46] K. Lerman and T. Hogg. Using a model of social dynamics topredict popularity of news. In
WWW , 2010.[47] J. Leskovec, L. Backstrom, and J. Kleinberg. Meme-trackingand the dynamics of the news cycle. In
KDD , 2009.[48] L. Leung. Generational differences in content generation insocial media: The roles of the gratifications sought and of nar-cissism.
Computers in Human Behavior , 29(3), 2013.[49] A. Liaw, M. Wiener, et al. Classification and regression byrandomForest.
R news , 2(3), 2002.[50] M. A. Lui, T. B. Penney, and A. Schirmer. Emotion effects ontiming: attention versus pacemaker accounts.
PloS one , 6(7),2011.[51] A. Marwick and R. Lewis. Media manipulation and disinfor-mation online.
New York: Data & Society Research Institute ,2017.[52] A. H. Maslow. A theory of human motivation.
Psychological Review , 50(4), 1943.[53] F. J. Massey Jr. The Kolmogorov-Smirnov test for goodnessof fit.
Journal of the American statistical Association , 46(253),1951.[54] B. McNair.
Fake news: Falsehood, fabrication and fantasy injournalism . Routledge, 2017.[55] H. Miller Hillberg, Z. Levonian, D. Kluver, L. Terveen, andB. Hecht. What I See is What You Don’t Get: The Ef-fects of (Not) Seeing Emoji Rendering Differences across Plat-forms.
Proceedings of the ACM on Human-Computer Interac-tion , 2(CSCW), 2018.[56] K. M. Miltner and T. Highfield. Never gonna GIF you up:Analyzing the cultural significance of the animated GIF.
SocialMedia+ Society , 3(3), 2017.[57] A. J. Mottram and M. J. Fleming. Extraversion, impulsivity,and online group membership as predictors of problematic In-ternet use.
CyberPsychology & Behavior , 12(3), 2009.[58] D. Mulfari, A. Celesti, M. Fazio, M. Villari, and A. Puliafito.Using Google Cloud Vision in assistive technology scenarios.In
IEEE ISCC , 2016.[59] A. Nagle.
Kill all normies: Online culture wars from 4chanand Tumblr to Trump and the alt-right . John Hunt Publishing,2017.[60] D. L. Nelson, V. S. Reed, and J. R. Walling. Pictorial supe-riority effect.
Journal of Experimental Psychology: Humanlearning and memory , 2(5), 1976.[61] K. Nelson-Field, E. Riebe, and K. Newstead. The emotionsthat drive viral video.
Australasian Marketing Journal , 21(4),2013.[62] K. Nelson-Field, E. Riebe, and B. Sharp. More mutterabout clutter: Extending empirical generalizations to Face-book.
Journal of Advertising Research , 53(2), 2013.[63] J. E. Opfer and S. A. Gelman. Development of the animate-inanimate distinction.
The Wiley-Blackwell handbook of child-hood cognitive development , 2, 2011.[64] B. Paris and J. Donovan. Deepfakes and Cheap Fakes.
Data &Society , 2019.[65] D. Parkhurst and E. Niebur. Scene content selected by activevision.
Spatial vision , 16(2), 2003.[66] R. Pieters and M. Wedel. Attention capture and transfer inadvertising: Brand, pictorial, and text-size effects.
Journal ofMarketing , 68(2), 2004.[67] T. E. Ruggiero. Uses and gratifications theory in the 21st cen-tury.
Mass Communication & Society , 3(1), 2000.[68] S. R. Safavian and D. Landgrebe. A survey of decision treeclassifier methodology.
IEEE transactions on systems, man,and cybernetics , 21(3), 1991.[69] M. J. Salganik, P. S. Dodds, and D. J. Watts. Experimentalstudy of inequality and unpredictability in an artificial culturalmarket.
Science , 311(5762), 2006.[70] C. Shao, G. L. Ciampaglia, A. Flammini, and F. Menczer.Hoaxy: A platform for tracking online misinformation. In
WWW Companion , 2016.[71] S. Shatford. Analyzing the subject of a picture: a theoreticalapproach.
Cataloging & Classification Quarterly , 6(3), 1986.[72] J. Z. Sojka and J. L. Giese. Communicating through picturesand words: Understanding the role of affect and cognition inprocessing visual and verbal information.
Psychology & Mar- eting , 23(12), 2006.[73] S. Sprecher, S. Treger, J. D. Wondra, N. Hilaire, and K. Wallpe.Taking turns: Reciprocal self-disclosure promotes liking in ini-tial interactions. Journal of Experimental Social Psychology ,49(5), 2013.[74] N. B. Stutts and R. T. Barker. The use of narrative paradigmtheory in assessing audience value conflict in image advertis-ing.
Management Communication Quarterly , 13(2), 1999.[75] B. Suh, L. Hong, P. Pirolli, and E. H. Chi. Want to beretweeted? large scale analytics on factors impacting retweetin twitter network. In
IEEE SCSM , 2010.[76] J. A. Suykens and J. Vandewalle. Least squares support vectormachine classifiers.
Neural processing letters , 9(3), 1999.[77] C. Tang, D. Garreau, and U. von Luxburg. When Do RandomForests Fail? In
NeurIPS , 2018.[78] B. W. Tatler. The central fixation bias in scene viewing: Select-ing an optimal viewing position independently of motor biasesand image feature distributions.
Journal of vision , 7(14), 2007.[79] I. Tsingalis, N. Vretos, N. Nikolaidis, and I. Pitas. Svm-basedshot type classification of movie content. In
MediterraneanElectrotechnical Conference , 2012.[80] L. Van Meurs and M. Aristoff. Split-second recognition: Whatmakes outdoor advertising work?
Journal of Advertising Re-search
IEEE Transactions on Pattern Analysis and MachineIntelligence , 39(4), 2016.[83] P. Vuilleumier. Facial expression and selective attention.
Cur- rent Opinion in Psychiatry , 15(3), 2002.[84] K. Q. Weinberger, J. Blitzer, and L. K. Saul. Distance metriclearning for large margin nearest neighbor classification. In
NeurIPS , 2006.[85] L. Weng, A. Flammini, A. Vespignani, and F. Menczer. Com-petition among memes in a world with limited attention.
Sci-entific Reports , 2, 2012.[86] L. Weng, F. Menczer, and Y.-Y. Ahn. Virality prediction andcommunity structure in social networks.
Scientific Reports , 3,2013.[87] A. Williams and M. Dupuis. I Don’t Always Spread Disinfor-mation on the Web, but When I do I Like to Use Memes: AnExamination of Memes in the Spread of Disinformation. In
IMCIC , 2020.[88] C. K. Williams and D. Barber. Bayesian classification withGaussian processes.
IEEE Transactions on Pattern Analysisand Machine Intelligence , 20(12), 1998.[89] L. Xie, A. Natsev, J. R. Kender, M. Hill, and J. R. Smith. Visualmemes in social media: tracking real-world news in youtubevideos. In
ACM Multimedia .[90] L. Yang, T. Sun, M. Zhang, and Q. Mei. We know what @you
WWW ,2012.[91] S. Zannettou, B. Bradlyn, E. De Cristofaro, G. Stringhini, andJ. Blackburn. Characterizing the Use of Images by State-Sponsored Troll Accounts on Twitter. In
ICWSM , 2020.[92] S. Zannettou, T. Caulfield, J. Blackburn, E. De Cristofaro,M. Sirivianos, G. Stringhini, and G. Suarez-Tangil. On the ori-gins of memes by means of fringe web communities. In
ACMIMC , 2018.[93] R. Zhou, J. Hentschel, and N. Kumar. Goodbye text, helloemoji: Mobile communication on WeChat in China. In
ACMCHI , 2017., 2017.