ArticlePDF Available

What Makes Politicians' Instagram Posts Popular? Analyzing Social Media Strategies of Candidates and Office Holders with Computer Vision

Authors:

Abstract and Figures

Previous research on the success of politicians' messages on social media has so far focused on a limited number of platforms, especially Facebook and Twitter, and predominately studied the effects of textual content. This research reported here applies computer vision analysis to a total of 59,020 image posts published by 172 Instagram accounts of U.S. politicians, both candidates and office holders, and examines how visual attributes influence audience engagement such as likes and comments. In particular, this study introduces an unsupervised approach that combines transfer learning and clustering techniques to discover hidden categories from large-scale visual data. The results reveal that different self-personalization strategies in visual media, for example, images featuring politicians in private, nonpolitical settings, showing faces, and displaying emotions, generally increase audience engagement. Yet, a significant portion of politician's Instagram posts still fell into the traditional, "politics-as-usual" type of political communication, showing professional settings and activities. The analysis explains how self-personalization is embodied in specific visual portrayals and how different self-presentation strategies affect audience engagement on a popular but less studied social media platform.
Content may be subject to copyright.
https://doi.org/10.1177/1940161220964769
The International Journal of Press/Politics
1 –24
© The Author(s) 2020
Article reuse guidelines:
sagepub.com/journals-permissions
DOI: 10.1177/1940161220964769
journals.sagepub.com/home/hij
Article
What Makes Politicians’
Instagram Posts Popular?
Analyzing Social Media
Strategies of Candidates and
Office Holders with Computer
Vision
Yilang Peng1
Abstract
Previous research on the success of politicians’ messages on social media has so
far focused on a limited number of platforms, especially Facebook and Twitter, and
predominately studied the effects of textual content. This research reported here
applies computer vision analysis to a total of 59,020 image posts published by 172
Instagram accounts of U.S. politicians, both candidates and office holders, and examines
how visual attributes influence audience engagement such as likes and comments. In
particular, this study introduces an unsupervised approach that combines transfer
learning and clustering techniques to discover hidden categories from large-scale
visual data. The results reveal that different self-personalization strategies in visual
media, for example, images featuring politicians in private, nonpolitical settings,
showing faces, and displaying emotions, generally increase audience engagement.
Yet, a significant portion of politician’s Instagram posts still fell into the traditional,
“politics-as-usual” type of political communication, showing professional settings and
activities. The analysis explains how self-personalization is embodied in specific visual
portrayals and how different self-presentation strategies affect audience engagement
on a popular but less studied social media platform.
Keywords
personalization, audience engagement, Instagram, virality, computer vision, image
clustering, unsupervised learning
1University of Georgia, Athens, GA, USA
Corresponding Author:
Yilang Peng, Department of Financial Planning, Housing and Consumer Economics, University of Georgia,
120 Barrow Hall, Athens, GA 30602, USA.
Email: yilang.peng@uga.edu
964769HIJXXX10.1177/1940161220964769The International Journal of Press/PoliticsPeng
research-article2020
2 The International Journal of Press/Politics 00(0)
While traditional political communication is often mediated by news coverage, social
media platforms allow politicians to circumvent news gatekeepers to raise their pro-
files and directly communicate with citizens. As a consequence, there is growing
interest in studying politicians’ messages on social media and what makes them
effective (Bene 2017; J. Lee and Xu 2018; McGregor et al. 2017; Metz et al. 2020;
Rossini et al. 2018). Yet, this line of scholarship has so far focused on a limited num-
ber of platforms, especially Facebook and Twitter, while political communication is
also vibrant on other social media platforms such as Instagram (Larsson 2019;
O’Connell 2018). With 37 percent of U.S. adults reporting use of this site, Instagram
has become the third most popular social media platform in the United States, follow-
ing YouTube and Facebook.1
This research applies computer-assisted image analysis, or computer vision, to a
total of 59,020 posts published by 172 Instagram accounts of U.S. politicians, includ-
ing candidates and office holders, to examine how visual attributes influence audience
engagement in the form of likes and comments. This study aims to advance our theo-
retical understanding of political communication on social media in several ways.
First, previous studies tend to focus on how textual attributes influence the success of
politicians’ messages on social media (e.g., J. Lee and Xu 2018). Some studies have
demonstrated the differential effects of using multimodal presentations—texts and
visual elements—on user reactions (Bene 2017; Casas and Williams 2019; J. Lee and
Xu 2018), yet the effects of specific visual attributes are less understood.
Research has also demonstrated the effectiveness of visuals in shaping viewers’
judgment of politicians (Grabe and Bucy 2009; Joo et al. 2014; Peng 2018; Shah
et al. 2016; Shah et al. 2015). There is a gap between the abundance of visuals in
politicians’ messages on social media and our understanding of how effectively dif-
ferent political messages influence audience engagement, however. To cite just one
example, a common strategy on political social media is for candidates to personal-
ize their campaigns, and empirical studies support the idea that self-personalization
helps politicians gain favorable impressions (E.-J. Lee and Oh 2012; McGregor
2018; Meeks 2017). Yet there is a dearth of studies on the visual aspect of personal-
ization. This study consequently engages with the concept of personalization and
specifically examines how it is reflected in the visual representations of candidates
and office holders.
Although the analysis of visual media is of theoretical importance, on a large
scale it does pose methodological challenges to communication scholars. Researchers
are now equipped with a wide range of natural language processing tools for analyz-
ing textual data (Manikonda et al. 2016; McGregor et al. 2017; Peng 2017; Rossini
et al. 2018). The field of computer vision, which trains computers to detect and
recognize specific digital imagery, has also provided social scientists with potential
tools of visual analysis (Joo et al. 2019; Peng 2018; Webb Williams et al. 2020). In
particular, while scholars can utilize unsupervised learning tools such as topic mod-
eling to discover latent patterns in a corpus of textual data, similar tools can also be
useful for communication researchers to find potential categories in a large amount
of visual media. This paper introduces an unsupervised approach that combines
Peng 3
transfer learning and clustering to detect “visual topics” from digital images, in the
hopes of enabling communication researchers to extract meaningful and interpreta-
ble patterns from the abundance of visual data that proliferates in today’s political
media environment.
Personalization as a Multifaceted Concept
To provide a context for the analysis of candidate visual self-representations, this
study highlights the concept of personalization, which has received growing attention
in the scholarship of political social media strategy (Bene 2017; Larsson 2019; Meeks
2016; Metz et al. 2020). Scholars generally see personalization as a multifaceted con-
cept. One prevalent approach proposes two components of personalization: (1) indi-
vidualization, meaning that political communication increasingly focuses on specific
candidates or politicians rather than parties, other institutions, or issues; and (2) priva-
tization, which refers to the tendency of portrayals to depict politicians as private
individuals rather than representatives of the people in public roles (Adam and Maier
2010; Holtz-Bacha et al. 2014; Van Aelst et al. 2012). Following this distinction, com-
munications posted on politicians’ personal social media accounts can be considered a
form of individualized communication (Metz et al. 2020).
Beyond the individualization–privatization distinction, scholars have argued that
the disclosure and expression of emotions should be conceptualized as another impor-
tant element of personalization (Metz et al. 2020; Van Santen and Van Zoonen 2020).
The emphasis on emotionalization, which describes communications about political
events and experiences that are more distinctive for their emotionality than their policy
substance, is recognized by Van Santen and Van Zoonen (2020) in their framework for
assessing personalized political communications.
Personalization and Social Media Engagement
A rich line of scholarship has examined what characteristics make online content pop-
ular or viral (Berger 2014). Noting that there is no consensus on the definition of
“virality,” this study takes a holistic view of virality as broadly encompassing a variety
of online interaction behaviors such as viewing, liking, commenting, and sharing
(Alhabash and McAlister 2015). Research has demonstrated that content that is likely
to “go viral” and receive a substantial amount of engagement tends to provoke emo-
tional arousal, contain information utility, and possess a novel or controversial quality
(Berger 2014; Casas and Williams 2019). Interestingly, viral messages could be posi-
tive or negatively valenced, depending on the content domain.
Previous research on personalization sometimes assumes that personalized media
coverage might disfavor political candidates, as this type of coverage might trivialize
office seekers and render them irrelevant (Aday and Devitt 2001). However, empirical
studies generally support the finding that self-personalization (and in particular, priva-
tization) helps politicians emotionally connect with viewers and foster favorable
impressions (Colliander et al. 2017; E.-J. Lee and Oh 2012; McGregor 2018; Meeks
4 The International Journal of Press/Politics 00(0)
2017). In experimental research, participants exposed to personalized tweets from
candidates were more likely to report feeling a sense of parasocial interaction and
social presence than those exposed to issue-oriented tweets alone (E.-J. Lee and Oh
2012; McGregor 2018). Viewers also perceived candidates who were presented in
personalized terms on Twitter as more likable and capable of dealing with political
issues than depersonalized candidates (Meeks 2017).
Analyses of social media data have shown that politicians’ personalized posts gar-
ner more audience feedback in the form of likes and comments than depersonalized
communications (Bene 2017; Larsson 2019; Metz et al. 2020). Among other reasons
for this is that personalized content tends to be emotional, a content feature that drives
virality (Berger 2014). In the Hungarian election of 2014, candidates’ Facebook posts
about family members attracted more likes and comments than posts about policy
(Bene 2017). From an analysis of German parliament members’ Facebook accounts,
Metz et al. (2020) showed that posts disclosing politicians’ private lives or emotional
side induced more audience expressions of “sentiments” (i.e., likes and emojis) than
depersonalized posts. Instagram, the focus of the present study, seems built for sharing
personalized content. In an examination of posts generated by users of both Instagram
and Twitter, Manikonda et al. (2016) showed that posts about nonpolitical content
such as art, food, fitness, fashion, travel, friends, and family, prevailed on Instagram,
whereas posts related to news, sports, and business were more popular on Twitter.
Despite the potential positive effects of personalization, research has documented
that politicians still use social media in a traditional, “politics-as-usual” manner, that
is, with a low level of personalization (Bene 2017; McGregor et al. 2017; Meeks
2016). Bene (2017) found that just 4 percent of candidates’ Facebook posts during the
2014 Hungarian election could be classified as personal. McGregor et al. (2017)
examined U.S. gubernatorial candidates that same year and revealed that only 7.3
percent of posts on Twitter and 9 percent on Facebook were personalized. Meeks
(2016) analyzed U.S. Senate candidates’ Twitter feeds in the 2012 election and found
that 11.8 percent of the candidate tweets included elements of personalization.
Personalization in Visual Media
How does personalization manifest itself in visual posts on Instagram? To answer this
question, we first review research on personalization in textual contexts. From an anal-
ysis of campaign websites, Hermans and Vergeer (2013) identified three dimensions or
foci of personalization: professional careers (e.g., official positions, political achieve-
ments), personal preferences (e.g., favorite music or sports), and family (e.g., marital
status, children). Trimble et al. (2013) considered demographic and socially salient
indicators of personalization, including gender, age, physical appearance, sexual iden-
tity/behavior, upbringing, marital status, and children.
In one of the few studies that systematically examined politicians’ Instagram posts,
O’Connell (2018) found that among the Instagram posts of members of Congress in
the United States during the first six months of 2017, just 8.16 percent were catego-
rized as personal, with the majority of posts (69.37%) classified as professional. Other
Peng 5
types of posts also emerged, such as text description of political statements and issue
positions (10.17%) and landscape photos featuring natural scenery without people
(2.84%) (O’Connell 2018). Larsson (2019) analyzed the top 20 most liked and com-
mented on Instagram posts from Norwegian party leaders and observed many posts
reflect politicians’ private lives. However, the analysis of personalization in visual
media remains limited. Therefore, the current study adopted an unsupervised machine
learning approach by first identifying the types of visual categories in politicians’
Instagram posts. The first research question therefore asks:
Research Question 1 (RQ1): What visual categories are presented in politicians’
image posts on Instagram?
Given that previous research has found a positive influence of privatized commu-
nications on audience engagement, we would expect a similar outcome for visual rep-
resentations of privatization. Therefore, the first hypothesis predicts that
Hypothesis 1 (H1): Among the categories identified in RQ1, visual communica-
tions related to the nonpolitical, private lives of politicians are more likely to garner
audience engagement (i.e., likes and comments) than communications related to
the professional and political lives of politicians.
In the context of visual communication, showing faces might be a simple strategy
of personalizing and cultivating close relationships with followers on social media.
This research situates face disclosure as an aspect of personalization in visual media
that is independent of privatization, as politicians can choose (or not) to show their
faces both in professional and private settings. The strategy of showing faces should
be particularly effective on a platform like Instagram, which is characterized by a
prevalence of selfies (Deeb-Swihart et al. 2017). The act of self-disclosure, revealing
personal information about oneself, often makes a person more liked by others (Collins
and Miller 1994). Studies in computer-mediated communication also demonstrate that
disclosing profile images fosters more favorable impressions among viewers (Feng
et al. 2016).
Since humans are evolutionarily drawn to human faces (Valenza et al. 1996), it is
possible that faces on social media will also be attention-getting (Bakhshi et al. 2014).
Prior research has shown that on Instagram, pictures with faces are more likely to get
likes and comments than posts without faces (Bakhshi et al. 2014). It is unclear, how-
ever, whether this tendency was due to people recognizing and engaging with faces
they know or whether users are attracted to faces in general. Therefore, this study aims
to isolate the effects of a politician’s own face and contrast it with both the absence of
any faces and the faces of others. We expect that images that foreground a politician’s
own face should get the highest engagement. We therefore predict that
Hypothesis 2 (H2): Images with politicians’ faces should attract more audience
engagement than images without politicians’ faces. Specifically, images with
6 The International Journal of Press/Politics 00(0)
politicians’ own faces should get more engagement than images without faces
(H2a) and images featuring only other people’s faces (H2b).
Related to this, we expect that the relative size of a politician’s face on the screen
should also be positively associated with increased engagement, based on the ten-
dency for larger objects to have higher visual saliency and a greater likelihood of
attracting viewer attention (Proulx 2010). Therefore,
Hypothesis 3 (H3): The area that a politician’s face occupies in an Instagram image
should be positively associated with greater audience engagement.
This study also pays attention to politicians’ facial expressions, which serve as
important nonverbal cues (Grabe and Bucy 2009). Consistent with the argument that
the expression of emotions should be conceptualized as an important element of per-
sonalization, Metz et al. (2020) found that Facebook posts published by politicians
were more likely to invite expressed sentiments (likes and emojis) and comments if
they involved emotional expressions of politicians or content with emotional appeal.
Research has shown that emotional arousal, regardless of its positive or negative
valence, drives content virality in general (Berger 2014). We thus expect that politi-
cians’ expressions of emotion should increase audience engagement. But does the
direction of expressed emotion matter?
Indeed, it does. A recurring finding in the viral media literature is that negativity works
effectively in propagating political messages on social media, especially Twitter (J. Lee
and Xu 2018; Stromer-Galley et al. 2018). Analyzing tweets from candidates competing
for U.S. governorships in 2014, Stromer-Galley et al. (2018) showed that candidates’
attack messages got retweeted more than their advocacy tweets. J. Lee and Xu (2018)
showed about half of Clinton and Trump’s tweets in the 2016 presidential campaign were
attack messages, which invited more retweets and likes than other types of messages.
While Twitter has developed a reputation for political hostility and negativity,
Instagram has yet to fall into this bitter pattern and, as a primarily visual platform,
might be more oriented toward positive self-presentation and sustaining social rela-
tionships than division. Some research confirms that the platform used drives the asso-
ciated tone of communication. Manikonda et al. (2016) showed that the same group of
users reported expressing more negative emotions and use of more work-related and
swear words on Twitter, while reporting use of more social words related to home,
family, and friends on Instagram.
During campaigns, displays of happiness typically make politicians look friendlier
and more competent, leaving positive impressions among viewers (Joo et al. 2014;
Peng 2018). We therefore expect that displays of happiness will also be associated
with increased audience engagement. Yet, given some emerging platform-oriented
norms, it is unclear if showing negative portrayals and expressions of emotion will be
effective in driving engagement on Instagram, a platform that emphasizes positive
self-presentations and social relationships. To explore these relationships, we propose
a hypothesis and research question:
Peng 7
Hypothesis 4 (H4): Politicians’ expression of positive emotion will be positively
associated with increased audience engagement on Instagram.
Research Question 2 (RQ2): How does the expression of negative emotion by
politicians influence audience engagement on Instagram?
Computer Vision Methods in Computational
Communication Research
In addition to documenting specific social media dynamics related to political messag-
ing on Instagram, this paper aims to make a methodological contribution by exploring
the potential of computer vision methods in analyzing political visuals. In the last
decade, deep neural networks (DNNs) have become a popular technique for many
computer vision tasks such as image classification. A DNN comprises an assemblage
of connected neurons organized in multiple layers. In a feedforward DNN, each neu-
ron receives inputs from neurons in the previous layer, performs computations, and
passes outputs to connected neurons in the subsequent layer (Chollet 2017; Joo and
Steinert-Threlkeld 2018). A convolutional neural network (CNN) is a specific cate-
gory of DNN that has several types of layers, such as convolutional layers and max-
pooling layers, that can handle the high dimensionality of image data (Chollet 2017;
Joo and Steinert-Threlkeld 2018). A CNN learns local patterns from raw pixels of
images. These patterns are (1) translation invariant, meaning that after a CNN learns a
certain pattern in one part of an image, it can identify it in other places of the image,
and (2) spatially hierarchical, meaning that after one CNN layer learns some small
local patterns such as edges and color patches, the following layers can recognize
more complicated and advanced patterns based on these simple patterns (Chollet
2017). These more complicated features are then used to recognize more complex and
abstract visual concepts (Chollet 2017).
A review of the existing literature suggests that there are at least four approaches
from the field of computer vision that can aid social scientists’ endeavors in visual
analysis. First, a wide range of open-source computer vision libraries (e.g., OpenFace)
and commercial APIs (e.g., Face++, Microsoft Azure) can help researchers perform
standardized and commonly used tasks, such as facial detection, emotion analysis,
object detection, and optical character recognition (e.g., Peng 2018). Researchers can
conveniently run analyses for a given input image, although they have to rely on a
limited number of visual attributes preselected by these libraries or services.
Researchers can also build prediction models for other outcomes using features pro-
vided by these tools (Joo et al. 2019).
Second, for supervised learning tasks, researchers can also prepare visual attributes
by themselves and train a neural network to predict the labels (Joo and Steinert-
Threlkeld 2018; Webb Williams et al. 2020). Typically, neural networks are trained on
a large number, sometimes millions, of labeled images. With relatively small data sets,
typically found in social sciences, researchers can resort to transfer learning tech-
niques, that is, extracting features from (or fine-tuning) a pre-trained network that has
been previously trained on a large data set (Chollet 2017). As noted, CNNs are able to
8 The International Journal of Press/Politics 00(0)
recognize advanced and complicated features. These features are originally used to
classify images in one training context, but they can be repurposed for a different task.
By training a classifier with advanced features from pre-trained models and new
labels, instead of training a brand-new model from raw pixels of images, researchers
can achieve good accuracy with relatively small data sets.
Transfer learning has commonly been used for supervised learning, but scholars
have also noted that this method can also be applied to unsupervised learning, for
example, grouping objects in pictures into similar-looking categories (Guérin et al.
2017). Unlike supervised learning, unsupervised learning uncovers hidden patterns in
data without pre-existing labels. The research reported here details a procedure of
combining transfer learning (specifically, feature extraction) and clustering to detect
visual topics from politicians’ images.
Last, visual messages can not only be described in terms of content but also in
terms of aesthetics. Scholars can now computationally calculate a variety of aesthetic
features of images, including brightness, blur, color, and composition. Prior work has
demonstrated associations between these visual attributes and outcomes such as image
virality and aesthetic appeal (Ke et al. 2006; Peng and Jemmott 2018). Recent works
have also started to harness the power of neural networks to assess the aesthetic appeal
of images (Talebi and Milanfar 2018).
Method
Sample
The study sample, which was curated from a list of U.S. politicians in autumn 2018,
included four groups: (1) candidates from the 2016 U.S. presidential primaries (N =
29), including seventeen Republican, six Democratic, four third-party candidates
(Gary Johnson, Jill Stein, Evan McMullin, and Darrell Castle), and two vice presiden-
tial candidates (Mike Pence, Tim Kaine); (2) governors (N = 50); (3) senators in the
115th Congress (N = 104), including four who ended their terms earlier; and (4) cabi-
net members of the Trump administration (N = 30). Some politicians fell into multiple
categories (e.g., Bernie Sanders). Duplicates were removed and a check was per-
formed to make sure each politician in the database had an Instagram account. Private
accounts and accounts with fewer than 20 posts were excluded. In some cases, a politi-
cian might have multiple accounts—these were kept in the sample. Altogether, 176
accounts representing 159 politicians (women = 19.5%, Democrat = 39.6%,
Republican = 56.6%, mean age = 62.4), were included in the analysis.
This study retrieved each politician’s entire Instagram feed published before 31
August 2018, along with each post’s caption, publication date, and numbers of likes
and comments.2 Only publicly accessible posts from politicians were downloaded.
The data collection occurred during the first week of September 2018. Videos were
excluded from the data set. Regarding “carousel” posts, a type of Instagram post that
contained multiple photos for viewers to swipe through, only the cover image was
kept. The final sample included 59,020 images. Each account had 335.3 image posts
on average (SD = 415.7, Mdn = 219.5).
Peng 9
Visual Categories
The analysis adopted an unsupervised approach that combined transfer learning and
clustering to categorize a large amount of visual data (Guérin et al. 2017). This pro-
cedure first converted each image to a vector of features with a pre-trained neural
network. To decide which pre-trained model to use for feature extraction, it is impor-
tant to consider the similarity between the data on which a model has been trained
and the data to be analyzed. Many publicly available pre-trained models are trained
on the ImageNet Large Scale Visual Recognition Challenge data set, which contains
approximately 1.4 million labeled images grouped into 1,000 categories (Chollet
2017). The ImageNet data set uses categories mostly related to animals (e.g., sting-
ray) and everyday objects (e.g., joystick, volleyball) and is frequently used for a
general-purpose image classification task.
However, regarding politicians’ images, one crucial aspect that distinguishes pro-
fessional and personal photos is the setting in which politicians present themselves.
The analysis thus integrated a data set about scene categorization, the Places365 data
set, which contains about 1.8 million images of 365 scene categories (Zhou et al.
2017). Some scene categories in the Places365 data set clearly related to politicians’
professional lives, such as office, legislative chamber, and conference center, whereas
other categories like kitchen and soccer field were more reflective of politicians’ pri-
vate lives. Therefore, the analysis utilized a VGG16 model pre-trained on a combina-
tion of the ImageNet and Places365 data sets.3
Next, the analysis fed each image into this pre-trained model and extracted features
from the third to last layer (fc2). The extracted features had 4,096 dimensions. A prin-
cipal component analysis (PCA) was applied on these dimensions with scikit-learn, a
machine learning library in Python. The first 200 factors (explaining 64.6% of the
variance) were used in clustering.
The analysis then applied a commonly used clustering algorithm k-means. One
challenge in k-means clustering is to choose the optimal number of clusters. This
research adopted an approach that kept human interpretation in the loop. First, the
model was repeatedly run with the number of clusters ranging from five to fifteen.
Second, in each clustering solution, twenty images were randomly selected for inspec-
tion from each cluster and exploratorily examined to confirm whether the images clas-
sified into each cluster formulated a coherent category. This step of inspection
identified four broad categories in the data set (see Table 1).
The first category, labeled “professional/political setting,” featured people in pro-
fessional or political settings, such as governmental buildings, legislative chambers,
press conferences, offices, conference rooms, rallies, and protests. The second cate-
gory, labeled “text/illustration,” was comprised mostly of textual messages, figures,
and illustrations. Politicians published this type of post to get their messages out,
including policy positions, event announcements, and supporter mobilization efforts.
These two categories were more related to the professional side of holding or seeking
political office. The third category, “personal setting,” showed individuals in private or
nonpolitical settings, such as bars, restaurants, shops, gyms, homes, vehicles, streets,
nature scenes, and settings that did not have clear visual cues related to politics (e.g.,
10
Table 1. Examples of Images in Each Category.
Category URL Account Short Description
Professional/
political setting
https://www.instagram.com/p/lNR3qlyqwf/ philbryantms Phil Bryant speaking to an audience in a
chamber.
https://instagram.com/p/mYZd1kubsZ/ kirstengillibrand Kirsten Gillibrand participating in a panel
discussion at a conference.
https://www.instagram.com/p/1_5sCfxiOA/ massgovernor Charlie Baker signing an executive order.
Text/illustration https://www.instagram.com/p/7yBMaPrJuw/ senjoniernst Joni Ernst sharing a picture of the POW/
MIA Flag.
https://www.instagram.com/p/
BVAf17rgYwG/
realdonaldtrump A quote from Donald Trump criticizing
the “FAKE MSM.”
https://www.instagram.com/p/
BYRSRV3D-P2/
kamalaharris Kamala Harris posting a message about
Women’s Equality Day.
Personal setting https://www.instagram.com/p/BUFesB4jGcC/ senatormenendez Bob Menendez with his mother.
https://www.instagram.com/p/Bgw-zhQBzgz/ nikkihaley Nikki Haley with her dog.
https://instagram.com/p/33npEfhPGu/ govmikehuckabee Mike Huckabee chatting with people in a
restaurant.
Architecture/
landscape
https://www.instagram.com/p/BeHE18znfQh/ Senatormartinheinrich Martin Heinrich sharing a photo of White
Sands National Monument.
https://www.instagram.com/p/
BHbClW7hJoh/
nygovcuomo Andrew M. Cuomo sharing a photo of
One World Trade Center.
https://www.instagram.com/p/7_Go_rkEyF/ senatortimscott Tim Scott sharing a photo of Lake
Greenwood.
Peng 11
a blurred background or wall). The fourth category, “architecture/landscape,” featured
views of buildings, structures, and landscapes, often in the absence of people. These
latter two categories were more related to the nonpolitical or private side of
politicians.
Next, each cluster in each solution was then assigned to one of the four categories
based on the twenty images selected from each cluster. This step applied the assigned
category to all of the images within that cluster. For example, if one cluster was
assigned to the “architecture/landscape” category, the label “architecture/landscape”
was then assumed to apply to every image in that cluster. This step of analysis also
suggested that when the number of clusters was too many (e.g., more than twelve
clusters), some clusters did not have the majority of images (>50%) clearly fitting into
one category. These solutions were removed from consideration.4
Two additional steps were taken to validate this process. To begin with, the agree-
ment between each pair of clustering solutions (5–11 clusters) regarding whether
images were consistently assigned to the same category was calculated. With the
exception of the five-cluster solution, the assigned categories based on different clus-
tering solutions were relatively consistent. There was generally over 80 or 90 percent
agreement among different clustering solutions (Figure 1) and the percentages of
images in each category in the data set were also similar (Figure 2), suggesting this
method classified images into the four categories in a relatively reliably way.
In addition, a total of six clustering solutions (6–11 clusters) were selected for fur-
ther validation. For each solution, twenty images were randomly inspected from the
four categories and checked for whether they matched their assigned categories.
Results suggested that in general, 75 to 95 percent of the images in each of the four
categories were classified correctly. The current study presents the results from the
Figure 1. Agreement on visual categories coded from different clustering solutions: (a)
percent agreement and (b) Cohen’s kappa.
12 The International Journal of Press/Politics 00(0)
11-cluster solution, in which 90, 80, 80, and 80 percent of the images in the four cat-
egories were identified correctly, respectively (see Supplementary Information file).
Facial Analysis
For facial analysis of the politician images, this study used Face++, a computer
vision API that specializes in facial detection and recognition. A target face set that
included faces of all the politicians sampled in this study was first prepared.5 For each
image in the data set, the facial recognition algorithm detected whether the image
featured a face or not. Next, for each detected face, the algorithm compared this face
to all the faces in the target face set and returned the most similar-looking face. The
majority (72.7%) of posts contained at least one face (M = 3.0, SD = 4.4). Each post
had on average 0.44 (SD = 0.57) faces belonging to the politician who owned the
account and 2.56 (SD = 4.19) faces of other people. A manual validation suggested
that the facial recognition algorithm accurately identified politicians’ faces, especially
when faces were near-frontal, not blocked, not blurry, and not too small (see
Supplementary Information file).
Based on the facial recognition results, the analysis compared four types of images
to isolate the effects of a politician’s own face compared to other people’s faces:
images with no face (27.3%), images with only the politician’s face (7.9%), images
with only faces of other people (31.7%), images with both the politician’s face and
other people’s faces (33.0%).
For politicians’ face size and emotional expression, the analysis only included
images (N = 24,190, 41.0%) that contained at least one face of the politician. First,
Face++ provided the location of the face as a rectangle. The analysis calculated how
much area the politician’s face occupied in the whole image, measured as the ratio
between the size of the facial rectangle and the whole image (Peng 2018). Natural log
Figure 2. Percentages of images assigned to the four categories based on different
clustering solutions.
Peng 13
transformation was applied for easier interpretation. In addition, Face++ also pro-
vided the probability of occurrence for seven discrete emotional expressions (i.e.,
happiness, anger, sadness, disgust, surprise, fear, and neutral emotion) ranging from
0 to 1. The analysis included the expression of happiness (M = 0.64, SD = 0.43) and
combined negative emotions (all other emotions detected except happiness and neu-
tral, M = 0.16, SD = 0.29) in the politician’ faces. Since all the emotions added to a
constant of one, neutral expressions were treated as the baseline for comparison. For
pictures featuring multiple faces of the same politician (e.g., a photo collage), average
values were used.
Control Variables
The analysis controlled for (1) aesthetic features commonly found to influence image
popularity, including brightness, contrast, colorfulness, and visual complexity (see
Supplementary Information file); (2) the number of days between the day when the
account’s first post in the data set was published and the day when this particular post
was published; (3) characteristics of each post’s caption, including word count and
percentages of positive/negative emotion words (calculated by Linguistic Inquiry and
Word Count) as well as the numbers of mentions (@) and hashtags (#); (4) politician
characteristics, including gender, age, race, party affiliation, and political role (e.g.,
senator, governors, presidential candidate, cabinet member); and (5) account charac-
teristics, including the numbers of followers, accounts following, and posts.
Results
With an R package lme4, a series of multilevel regression analyses (N = 59,020)
examined what visual features predicted audience engagement for each image post
(Table 2). The analyses treated accounts (N = 176) as random effects. Visual catego-
ries, facial features, and control variables mentioned earlier were included as fixed
effects. To facilitate statistical analysis, natural log transformation was applied to the
two highly skewed dependent variables, the numbers of likes (skew = 7.5) and com-
ments (skew = 230.1) a photo received.
Visual Categories
RQ1 asks what visual categories are present in politicians’ images. As noted, the clus-
tering method identified four broad categories (Table 1), including two related to the
professional and political side of politicians, professional/political setting (62.1%) and
text/illustration (12.3%), and two related to the personal side of politicians, personal
setting (16.1%) and architecture/landscape (9.5%).
The analysis then looked at how audiences liked different visual categories (H1).
The category professional/political setting was treated as the reference group. The
other politics-related category, text/illustration, did not significantly differ from the
14 The International Journal of Press/Politics 00(0)
reference category. In comparison, the personal setting category (b = 0.19, p < .001)
and the architecture/landscape category positively contributed to likes (b = 0.14,
p < .001). The outcome variable was log-transformed. Compared with images in the
reference category, images in these two categories attracted more likes by 21 and 15
percent, respectively.
The engagement pattern was slightly different for comments. Compared with the
professional setting category, the text/illustration category attracted more comments
by 13 percent (b = 0.12, p < .001). Regarding nonpolitical categories, the personal
setting category still invited a large number of comments with a 16 percent increase
Table 2. Multilevel Analyses Predicting the Numbers of Likes and Comments.
Likes (Log-Trans.) Comments (Log-Trans.)
b95% CI b95% CI
Estimates of fixed effects
Politician characteristics
Women 0.18 [−0.18, 0.54] 0.01 [−0.23, 0.25]
(ref. = Republican)
Democrat
−0.01 [−0.31, 0.28] −0.10 [−0.29, 0.10]
Independent 0.57 [−0.20, 1.35] 0.33 [−0.18, 0.84]
(ref. = mixeda)
Senate
−1.07*** [−1.64, −0.50] −0.75*** [−1.13, −0.38]
Governor −0.59 [−1.19, 0.01] −0.53* [−0.93, −0.14]
Cabinet −0.31 [−1.12, 0.49] −0.12 [−0.66, 0.41]
Presidential candidate 0.24 [−0.45, 0.92] 0.04 [−0.41, 0.50]
Nonwhite −0.40 [−1.00, 0.21] −0.44* [−0.84, −0.04]
Age 0.00 [−0.01, 0.02] 0.01 [0.00, 0.02]
Visual category (ref. = professional/political setting)
Text/illustration −0.01 [−0.03, 0.01] 0.12*** [0.10, 0.15]
Personal setting 0.19*** [0.17, 0.21] 0.15*** [0.13, 0.17]
Architecture/
landscape
0.14*** [0.12, 0.16] 0.00 [−0.03, 0.02]
Face category (ref. = no faces)
Politician’s face only 0.26*** [0.23, 0.28] 0.30*** [0.27, 0.33]
Other faces only 0.06*** [0.04, 0.08] 0.02* [0.00, 0.04]
Politician with other
faces
0.12*** [0.10, 0.13] 0.07*** [0.05, 0.09]
Standard deviations of random effects
Accounts (N = 176) 0.92 0.60
Residual 0.62 0.76
Note. N = 59,020. Unstandardized regression coefficients are shown. The full results including all the
control variables are in the Supplementary Information file (Table S1). CI = confidence interval.
a. Politicians belonging to multiple categories.
*p < .05. **p < .01. ***p < .001.
Peng 15
(b = 0.15, p < .001). But the architecture/landscape cluster, although aesthetically
appealing and likable, did not influence comments. Overall, the hypothesis (H1) that
nonpolitical, personal categories would receive more audience engagement was well-
supported regarding likes, but only received partial support regarding comments.
Consistent with H1, the personal setting category received more comments than the
professional setting category. Yet, the architecture/landscape category, another nonpo-
litical category, did not significantly differ from the professional setting category and
actually received fewer comments than the text/illustration category, which contra-
dicted H1.
Face Disclosure
To examine the effects of face disclosure, images without faces were treated as the
reference group. Compared to pictures with no faces, the other three categories, images
featuring only the politician (b = 0.26, p < .001), images featuring only other people
(b = 0.06, p < .001), and images featuring the politician with other people (b = 0.12,
p < .001), were all associated with an increase in likes. Images with only the politi-
cian’s face received the highest increase in likes (29%), followed by images featuring
the politician with other people (12%) and images featuring other people’s faces only
(6%).
A similar pattern occurred for comments. Images with faces received more com-
ments, with posts only featuring the politician’s face resulting in the highest increase
of 35 percent (b = 0.30, p < .001). Images featuring the politician with other people
and images with other faces were associated with a 7 percent (b = 0.07, p < .001) and
2 percent (b = 0.02, p = .045) increase in comments, respectively. In summary, faces,
in general, drove audience engagement. Images featuring politicians’ own faces only
were the most effective in spurring engagement, followed by images featuring politi-
cians with other faces, images containing only other faces, and images without faces.6
H2a and H2b were supported.
Face Size and Emotional Expressions
Last, the analysis investigated the effects of politicians’ face size and emotional expres-
sions. The following multilevel analyses only included images featuring politician’s
faces (N = 24,190; Table 3). First, compared with images featuring a politician’s face
only, images featuring both the politician and other faces received fewer likes by 10
percent (b = −0.11, p < .001) and fewer comments by 15 percent (b = −0.17,
p < .001), confirming the previous observation that the politician’s own face attracted
the most engagement. Face size positively contributed to both likes (b = 0.10,
p < .001) and comments (b = 0.11, p < .001), supporting H3. When face size doubled,
a post’s likes and comments increased by 7 and 8 percent, respectively.
Regarding emotional expressions, the expression of neutral emotion was treated as
the baseline for comparison. Facial displays of happiness positively influenced likes
(b = 0.06, p < .001): indeed, happy faces received 7 percent more likes than neutral
16 The International Journal of Press/Politics 00(0)
faces. However, happiness did not influence the number of comments. H4 thus
received partial support only regarding likes. Regarding RQ2, the expression of nega-
tive emotion increased a post’s likes by 6 percent (b = 0.06, p < .001), but did not
influence comments.
Discussion
In summary, this research advances our understanding of politicians’ use of social
media platforms, specifically Instagram, and how different self-personalization strate-
gies influence audience engagement. This study specifies different self-personaliza-
tion strategies in visual media, for example, appearing in private, nonpolitical settings,
showing one’s face, and displaying emotions, and empirically shows that these strate-
gies generally increase audience engagement. Beyond contributing to a theoretical
dialogue with the social media political communication literature, this study also illus-
trates the potential of applying computer vision techniques in political communication
research. In particular, this study describes an unsupervised approach to analysis that
combines transfer learning and clustering to extract “visual topics” from large-scale
visual media, which could be applied in future research.7
Table 3. Multilevel Analyses Predicting the Numbers of Likes and Comments (Only
Including Images Containing the Politician’s Face).
Likes (Log-Trans.) Comments (Log-Trans.)
b95% CI b95% CI
Estimates of Fixed Effects
Visual category (ref. = professional/political setting)
Text/illustration 0.09** [0.03, 0.15] 0.31*** [0.23, 0.38]
Personal setting 0.13*** [0.10, 0.15] 0.07*** [0.04, 0.11]
Architecture/
landscape
0.29*** [0.21, 0.36] 0.16*** [0.07, 0.25]
Face category (ref. = Politician’s face only)
Politician with other
faces
−0.11*** [−0.13, −0.09] −0.17*** [−0.19, −0.14]
Facial features
Face size (log-trans.) 0.10*** [0.09, 0.11] 0.11*** [0.10, 0.12]
Happiness 0.06*** [0.04, 0.09] 0.00 [−0.03, 0.03]
Negative emotions 0.06*** [0.02, 0.09] 0.04 [−0.01, 0.08]
Standard deviations of random effects
Accounts (N = 176) 0.90 0.60
Residual 0.60 0.76
Note. N = 24,190. Unstandardized regression coefficients are shown. The full results including all the
control variables are in the Supplementary Information file (Table S2). CI = confidence interval.
*p < .05. **p < .01. ***p < .001.
Peng 17
Personalization Strategies on Social Media
The analysis reported here examined the effects of different content categories on
audience engagement, with an emphasis on the role of privatization. A significant por-
tion of politician’s Instagram posts still fell into the traditional, “politics-as-usual”
type of political communication, showing professional activities, but this type of con-
tent was generally less successful in attracting engagement. Instead, nonpolitical and
private content from politicians generally attracted more engagement such as likes
and, less consistently, comments from audiences than political content. This result
echoes previous studies showing that politicians’ personalized content often receives
more audience reactions on Facebook (Bene 2017; Metz et al. 2020).
Due to the “social” nature of social media platforms, users might prefer more inti-
mate content from politicians and expect them to communicate and share aspects of
their daily lives more like an everyday person than like a politician (Bene 2017). The
preference for personalized content could also be due to a “novelty” factor: As politi-
cians’ Instagram feeds remain preoccupied with more traditional type of self-portray-
als, personalized pictures garner more likes and comments because they are still
somewhat unusual for politicians to share, thus receiving greater attention from
viewers.
This study also reveals that some content categories distinctly influence the volume
of likes and comments, indicating that social media users might have different psycho-
logical motivations when liking and commenting on politicians’ posts. Because politi-
cal visuals are more intuitive than textual messages and don’t require linguistic
parsing, liking an image might be less cognitively demanding than commenting on a
message (Alhabash and McAlister 2015). Liking also typically signals a positive eval-
uation of a media message while comments can be of mixed sentiment (Peng and
Jemmott 2018).
In the results, textual messages used by politicians, while not getting many likes,
led to substantially more comments. This revealed that texts added to visual images
could provoke more thoughts or cognitive efforts among viewers, reinforcing previous
findings from Facebook that text posts from politicians get more comments than likes
(Bene 2017). In comparison, images of architecture and landscape, while attracting a
large number of likes, failed to get many comments. This type of image might be aes-
thetically likable and serve as “eye candy” in politicians’ newsfeeds but does not pro-
voke conversations among followers.8
This study also examined the effectiveness of showing faces as a personalization
strategy in visual media. Prior research has shown that on social media, pictures with
faces are more likely to get likes and comments than pictures without faces (Bakhshi
et al. 2014). It is unknown, however, whether this tendency is due to the fact that
people engage with generic faces or that people are attracted to the faces of account
holders they are following. This research showed that although the presence of faces
did increase audience engagement overall, politicians’ own faces were the most effec-
tive in enhancing audience engagement. This finding indicates that social media users
18 The International Journal of Press/Politics 00(0)
react positively to images with faces primarily because they can recognize them and
build a kind of social connection.
In addition, larger faces also garnered more attention and engagement from audi-
ences. This finding is consistent with research on “face-ism,” which posits that the
prominence of faces in visual representations tend to make a person look more favor-
able (Archer et al. 1983), but might contradict some findings that large faces in media
outlets can elicit negative evaluation of politicians (Peng 2018). This could be due to
the fact that politicians’ Instagram posts are generally positive self-presentations and
do not feature extreme close-ups, which tend to be associated with negative portrayals
in news outlets (Grabe and Bucy 2009).
This study also shows the positive effect of displaying emotions as a personalizing
strategy on audience engagement. Prior research on virality suggests that content with
the capacity to provoke emotional arousal, regardless of valence, is likely to go viral
(Berger 2014). Similarly, the current study found that facial expressions of both posi-
tive and negative emotion increase likes for politicians at a higher rate than visual
portrayals showing neutral emotion. This pattern is also consistent with our observa-
tions regarding captions (see Supplementary Information file): posts with captions
containing either positively or negatively valenced words were associated with more
likes and comments. Therefore, although Instagram is generally oriented toward posi-
tive self-presentation and building social connections, users might still be drawn to
negative messages from politicians when it comes to politics.
Effects of Personalization beyond Audience Engagement
While this research reveals the effectiveness of personalization strategies in visual
content regarding audience engagement, future studies can investigate how personal-
ization influences user activity beyond social media engagement. Prior research has
shown that personalization can elicit more positive impressions and make viewers feel
more connected with a politician (E.-J. Lee and Oh 2012; McGregor 2018). By per-
sonalizing on social media, politicians might present themselves as relatable and
approachable, but this strategy might also render themselves as less serious contenders
and distract viewers from more substantial issues such as their policy positions.
Experimental studies have revealed an innuendo effect in person perception that
reveals the weaknesses of certain study designs: when media coverage of a politician
only focuses on one of the two trait dimensions, warmth and competence (i.e., only as
friendly or only as competent), viewers would rate the politician more negative on the
other dimension (Koch and Obermaier 2016). In addition, while personalization might
produce positive effects among less politically involved viewers, more politically
engaged voters could be less drawn to personalized coverage and prefer policy pro-
nouncements; hence, sharing images in this case could further develop negative feel-
ings toward politicians (E.-J. Lee and Oh 2012).
Nevertheless, politicians on social media platforms rarely use personalized content
alone. Consistent with prior research (O’Connell 2018), the results show that personal-
ized content is outweighed by more traditional, “politics-as-usual” posts. It is unlikely
Peng 19
that personalized posts on Instagram overshadow politicians’ delivery of policy stances
or display of more politically relevant qualities. Future studies can examine how a
combination of varying self-presentation strategies, which represents a more realistic
scenario on social media, affects the evaluation of politicians.
Limitations and Future Research
While this study provides valuable insights into Instagram, a popular but less exam-
ined platform (Larsson 2019; O’Connell 2018), the choice of examining this particular
platform might shape the observations of this research. The majority of Instagram
users do not turn to this site for political content. Only 32 percent of Instagram users
reported getting news on this platform, compared with 71 percent on Twitter and 67
percent on Facebook.9 This might contribute to the pattern of nonpolitical content
performing better in getting people’s attention. Regarding interaction behaviors,
Instagram prioritizes liking over other interaction behaviors such as reposting, which
might encourage more positive self-portrayals. Hence, a direction for future research
might be to look beyond a single social media platform and compare audience
responses to politicians’ self-presentations across different sites.
This research also shows that computer vision techniques can greatly expand the
scope of visual analysis in political communication and demonstrates the utility of an
unsupervised approach to categorizing political visuals. The proposed approach is not
without limitations. Certain images were assigned to the wrong categories. Some mis-
classifications could be due to the tendency of the clustering algorithm grouping visu-
ally similar-looking images based on transfer learning features without knowing the
actual content of the images. For example, some pictures of still objects such as a
medal were categorized as text/illustration, as these images often had words or pat-
terns that looked similar to captioned images.
Also, some images might be ambiguous and can be assigned into multiple catego-
ries. For example, an image showing a text statement on a natural scene background
might be assigned to both the text and the landscape category. If such images are
prevalent in the data set, it might be more appropriate to apply a mixed-membership
clustering method instead of the single-membership method used in this study,
k-means. In addition, although the pre-trained models used in this study were pri-
marily about everyday objects and scenes, it is unknown if the extracted features
might still incorporate gender, racial, or cultural biases, which should be further
investigated. Future research can experiment with different pre-trained models,
clustering algorithms, and analytical procedures to improve the performance of this
approach.
Future research can also apply computer vision methods to analyze political videos
on social media. Videos have been frequently used in political communication and
some studies have already applied computational methods in analyzing effects of
political videos (Joo et al. 2019; Shah et al. 2016; Shah et al. 2015). Combining com-
putational textual analysis and visual analysis is another promising research direction.
While this study only looked at the number of comments given to politicians’ posts,
20 The International Journal of Press/Politics 00(0)
analyzing social media users’ comments should give us more insights into viewers’
interpretations of these visual portrayals.
Acknowledgments
The author thanks Sandra González-Bailón, Michael X. Delli Carpini, Jessa Lingel, Yphtach
Lelkes, Diana C. Mutz, Tian Yang, audience of the 2019 Political Communication Workshop at
the University of Pennsylvania, two anonymous reviewers, and guest editors of the special issue
for their feedback on earlier versions of the study.
Declaration of Conflicting Interests
The author(s) declared no potential conflicts of interest with respect to the research, authorship,
and/or publication of this article.
Funding
The author(s) received no financial support for the research, authorship, and/or publication of
this article.
ORCID iD
Yilang Peng https://orcid.org/0000-0001-7711-9518
Supplemental Material
Supplemental material for this article is available online.
Notes
1. https://www.pewinternet.org/fact-sheet/social-media/
2. For data collection, this study used a program called 4K Stogram (https://www.4kdownload.
com/products/product-stogram) to download each account’s posts and a Python script
to retrieve each post’s meta data (e.g., caption, number of likes and comments) using
Selenium. For more discussion about the ethics of web scraping in computational com-
munication research, see Freelon (2018).
3. The pre-trained model (VGG16-hybrid1365) is available at https://github.com/
GKalliatakis/Keras-VGG16-places365.
4. The randomly selected images from each cluster, the assigned categories, and the randomly
selected images from each category are available at https://osf.io/9fmw5/
5. To prepare the target face set, Google image search was used to download a collection of
images of each politician. For each politician, the face set included three pictures that fea-
tured only one frontal face of the politician that was of good resolution and was in focus.
6. The differences among the four categories regarding likes and comments were all statisti-
cally significant, as indicated by additional regression analyses using the images with other
faces only and images featuring politicians with others as the reference groups.
7. Supporting data and Python codes can be found at https://osf.io/9fmw5/ and https://github.
com/yilangpeng/transfer-learning-clustering
8. As indicated by Tables 2 and 3, there might be interaction effects between visual catego-
ries and face categories. The analysis exploratorily tested this possibility by adding the
Peng 21
products of the three dummy variables for visual categories (text/illustration, personal
setting, and architecture/landscape) and the three dummy variables for face-related catego-
ries (politician only, other people only, both) in the regression models (see Supplementary
Information file Table S3). In general, the effects of showing the politician’s faces in
attracting likes and comments were weaker in the professional setting category than the
effects for the other three categories.
9. https://www.journalism.org/2018/09/10/news-use-across-social-media-platforms-2018/
References
Adam, Silke, and Michaela Maier. 2010. “Personalization of Politics a Critical Review and
Agenda for Research.” Annals of the International Communication Association 34 (1):
213–57.
Aday, Sean, and James Devitt. 2001. “Style over Substance: Newspaper Coverage of
Elizabeth Dole’s Presidential Bid.” Harvard International Journal of Press/Politics 6
(2): 52–73.
Alhabash, Saleem, and Anna R. McAlister. 2015. “Redefining Virality in Less Broad Strokes:
Predicting Viral Behavioral Intentions from Motivations and Uses of Facebook and
Twitter.” New Media & Society 17 (8): 1317–39.
Archer, Dane, Bonita Iritani, Debra D. Kimes, and Michael Barrios. 1983. “Face-ism: Five
Studies of Sex Differences in Facial Prominence.” Journal of Personality and Social
Psychology 45 (4): 725–35.
Bakhshi, Saeideh, David A. Shamma, and Eric Gilbert. 2014. “Faces Engage Us: Photos with
Faces Attract More likes and Comments on Instagram.” In Proceedings of the SIGCHI
Conference on Human Factors in Computing Systems. New York: Association for
Computing Machinery. doi:10.1145/2556288.2557403.
Bene, Márton. 2017. “Go Viral on the Facebook! Interactions between Candidates and Followers
on Facebook during the Hungarian General Election Campaign of 2014.” Information,
Communication & Society 20 (4): 513–29.
Berger, Jonah. 2014. “Word of Mouth and Interpersonal Communication: A Review and
Directions for Future Research.” Journal of Consumer Psychology 24 (4): 586–607.
Casas, Andreu, and Nora Webb Williams. 2019. “Images that Matter: Online Protests and the
Mobilizing Role of Pictures.” Political Research Quarterly 72 (2): 360–75.
Chollet, François. 2017. Deep Learning with Python. Shelter Island: Manning Publications.
Colliander, Jonas, Ben Marder, Lena Lid Falkman, Jenny Madestam, Erik Modig, and Sofie
Sagfossen. 2017. “The Social Media Balancing Act: Testing the Use of a Balanced Self-
Presentation Strategy for Politicians Using Twitter.” Computers in Human Behavior
74:277–85.
Collins, Nancy L., and Lynn Carol Miller. 1994. “Self-Disclosure and Liking: A Meta-Analytic
Review.” Psychological Bulletin 116 (3): 457–75.
Deeb-Swihart, Julia, Christopher Polack, Eric Gilbert, and Irfan Essa. 2017. “Selfie-Presentation
in Everyday Life: A Large-Scale Characterization of Selfie Contexts on Instagram.” In
Proceedings of the Eleventh International AAAI Conference on Web and Social Media.
Association for the Advancement of Artificial Intelligence. https://smartech.gatech.edu/
bitstream/handle/1853/60097/2017-Deeb-Swihart-SELLCSCI.pdf.
Feng, Bo, Siyue Li, and Na Li. 2016. “Is a Profile Worth a Thousand Words? How Online
Support-Seeker’s Profile Features May Influence the Quality of received support
Messages.” Communication Research 43 (2): 253–76.
22 The International Journal of Press/Politics 00(0)
Freelon, Deen. 2018. “Computational Research in the Post-API Age.” Political Communication
35 (4): 665–68.
Grabe, Maria Elizabeth, and Erik Page Bucy. 2009. Image Bite Politics: News and the Visual
Framing of Elections. Oxford: Oxford University Press.
Guérin, Joris, Olivier Gibaru, Stéphane Thiery, and Eric Nyiri. 2017. “CNN Features Are Also
Great at Unsupervised Classification.” https://arxiv.org/abs/1707.01700.
Hermans, Liesbeth, and Maurice Vergeer. 2013. “Personalization in E-Campaigning: A Cross-
National Comparison of Personalization Strategies Used on Candidate Websites of 17
Countries in EP Elections 2009.” New Media & Society 15 (1): 72–92.
Holtz-Bacha, Christina, Ana Ines Langer, and Susanne Merkle. 2014. “The Personalization
of Politics in Comparative Perspective: Campaign Coverage in Germany and the United
Kingdom.” European Journal of Communication 29 (2): 153–70.
Joo, Jungseock, Erik P. Bucy, and Claudia Seidel. 2019. “Automated Coding of Televised Leader
Displays: Detecting Nonverbal Political Behavior with Computer Vision.” International
Journal of Communication 13:4044–66.
Joo, Jungseock, Weixin Li, Francis F. Steen, and Song-Chun Zhu. 2014. “Visual Persuasion:
Inferring Communicative Intents of Images.” In Proceedings of the IEEE Conference
on Computer Vision and Pattern Recognition. Columbus: Institute of Electrical and
Electronics Engineers. https://www.cv-foundation.org/openaccess/content_cvpr_2014/
papers/Joo_Visual_Persuasion_Inferring_2014_CVPR_paper.pdf.
Joo, Jungseock, and Zachary C. Steinert-Threlkeld. 2018. “Image as Data: Automated Visual
Content Analysis for Political Science.” https://arxiv.org/abs/1810.01544.
Ke, Yan, Xiaoou Tang, and Feng Jing. 2006. “The Design of High-Level Features for Photo
Quality Assessment.” In Proceedings of the IEEE Computer Society Conference on
Computer Vision and Pattern Recognition. Washington, D.C.: Institute of Electrical and
Electronics Engineers. https://ieeexplore.ieee.org/document/1640788.
Koch, Thomas, and Magdalena Obermaier. 2016. “With Heart and (No) Mind? How Recipients
Negatively Infer Missing Information about Politicians and How This Affects the
Assessment of the Speaker.” Communication Research 43 (7): 972–95.
Larsson, Anders Olof. 2019. “Skiing All the Way to the Polls: Exploring the Popularity of
Personalized Posts on Political Instagram Accounts.” Convergence 25 (5–6): 1096–
110.
Lee, Eun-Ju, and Soo Youn Oh. 2012. “To Personalize or Depersonalize? When and How
Politicians’ Personalized Tweets Affect the Public’s Reactions.” Journal of Communication
62 (6): 932–49.
Lee, Jayeon, and Weiai Xu. 2018. “The More Attacks, the More Retweets: Trump’s and
Clinton’s Agenda Setting on Twitter.” Public Relations Review 44 (2): 201–13.
Manikonda, Lydia, Venkata Vamsikrishna Meduri, and Subbarao Kambhampati. 2016.
“Tweeting the Mind and Instagramming the Heart: Exploring Differentiated Content
Sharing on Social Media.” In Tenth International AAAI Conference on Web and Social
Media. Cologne: Association for the Advancement of Artificial Intelligence. https://arxiv.
org/abs/1603.02718.
McGregor, Shannon C. 2018. “Personalization, Social Media, and Voting: Effects of Candidate
Self-Personalization on Vote Intention.” New Media & Society 20 (3): 1139–60.
McGregor, Shannon C., Regina G. Lawrence, and Arielle Cardona. 2017. “Personalization,
Gender, and Social Media: Gubernatorial Candidates’ Social Media Strategies.”
Information, Communication & Society 20 (2): 264–83.
Peng 23
Meeks, Lindsey. 2016. “Gendered Styles, Gendered Differences: Candidates’ Use of
Personalization and Interactivity on Twitter.” Journal of Information Technology &
Politics 13 (4): 295–310.
Meeks, Lindsey. 2017. “Getting Personal: Effects of Twitter Personalization on Candiate
Evaluations.” Politics & Gender 13 (1): 1–25.
Metz, Manon, Sanne Kruikemeier, and Sophie Lecheler. 2020. “Personalization of Politics on
Facebook: Examining the Content and Effects of Professional, Emotional and Private Self-
Personalization.” Information, Communication & Society 23:1481–98.
O’Connell, David. 2018. “# Selfie: Instagram and the United States Congress.” Social Media+
Society 4 (4). doi:10.1177/2056305118813373.
Peng, Yilang. 2017. “Time Travel with One Click: Effects of Digital Filters on Perceptions
of Photographs.” In Proceedings of the 2017 CHI Conference on Human Factors in
Computing Systems, 6000–11. New York: Association for Computing Machinery.
doi:10.1145/3025453.3025810.
Peng, Yilang. 2018. “Same Candidates, Different Faces: Uncovering Media Bias in Visual
Portrayals of Presidential Candidates with Computer Vision.” Journal of Communication
68 (5): 920–41.
Peng, Yilang, and John B. Jemmott III. 2018. “Feast for the Eyes: Effects of Food Perceptions
and Computer Vision Features on Food Photo Popularity.” International Journal of
Communication 12:313–36.
Proulx, Michael J. 2010. “Size Matters: Large Objects Capture Attention in Visual Search.”
PLoS ONE 5 (12): e15293.
Rossini, Patrícia, Jennifer Stromer-Galley, Kate Kenski, Jeff Hemsley, Feifei Zhang, and
Brian Dobreski. 2018. “The Relationship between Race Competitiveness, Standing in
the Polls, and Social Media Communication Strategies during the 2014 US Gubernatorial
Campaigns.” Journal of Information Technology & Politics 15 (3): 245–61.
Shah, Dhavan V., Alex Hanna, Erik P. Bucy, David S. Lassen, Jack Van Thomme, Kristen Bialik,
JungHwan Yang, and Jon C. W. Pevehouse. 2016. “Dual Screening during Presidential
Debates: Political Nonverbals and the Volume and Valence of Online Expression.”
American Behavioral Scientist 60 (14): 1816–43.
Shah, Dhavan V., Alex Hanna, Erik P. Bucy, Chris Wells, and Vidal Quevedo. 2015. “The Power
of Television Images in a Social Media Age: Linking Biobehavioral and Computational
Approaches Via the Second Screen.” The ANNALS of the American Academy of Political
and Social Science 659 (1): 225–45.
Stromer-Galley, Jennifer, Feifei Zhang, Jeff Hemsley, and Sikana Tanupabrungsun. 2018.
“Tweeting the Attack: Predicting Gubernatorial Candidate Attack Messaging and Its
Spread.” International Journal of Communication 12:3511–32.
Talebi, Hossein, and Peyman Milanfar. 2018. “NIMA: Neural Image Assessment.” IEEE
Transactions on Image Processing 27 (8): 3998–4011.
Trimble, Linda, Angelia Wagner, Shannon Sampert, Daisy Raphael, and Bailey Gerrits. 2013.
“Is It Personal? Gendered Mediation in Newspaper Coverage of Canadian National Party
Leadership Contests, 1975–2012.” The International Journal of Press/Politics 18 (4): 462–81.
Valenza, Eloisa, Francesca Simion, Viola Macchi Cassia, and Carlo Umiltà. 1996. “Face
Preference at Birth.” Journal of Experimental Psychology: Human Perception and
Performance 22 (4): 892–903.
Van Aelst, Peter, Tamir Sheafer, and James Stanyer. 2012. “The Personalization of Mediated
Political Communication: A Review of Concepts, Operationalizations and Key Findings.”
Journalism 13 (2): 203–20.
24 The International Journal of Press/Politics 00(0)
Van Santen, Rosa, and Liesbet Van Zoonen. 2020. “The Personal in Political Television
Biographies.” Biography 33 (1): 46–67.
Webb Williams, Nora, Andreu Casas, and John Wilkerson. 2020. Images as Data for Social
Science Research: An Introduction to Convolutional Neural Nets for Image Classification.
Cambridge: Cambridge University Press.
Zhou, Bolei, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. 2017.
“Places: A 10 Million Image Database for Scene Recognition.” IEEE Transactions on
Pattern Analysis and Machine Intelligence 40 (6): 1452–64.
Author Biography
Yilang Peng (PhD, University of Pennsylvania) is an assistant professor in the Department of
Financial Planning, Housing and Consumer Economics at the University of Georgia. His
research is at the intersection of computational social science, visual communication, social
media analytics, and science communication.
... There is a growing work among communication scholars that analyzes visual frames in media messages (Bucy & Joo, 2021) and applies computational skills to study visual frames (Joo & Steinert-Threlkeld, 2018;Peng, 2021). These computational methods are applied to 1) extract high-level features such as capturing facial expressions in the images or videos, and (or) 2) pull out visual modalities such as color or saturation in the images or videos. ...
... Communication studies employing computational techniques to analyze visual frames follow at least four approaches (Peng, 2021): 1) extracting attributes predetermined by open-source computer vision libraries and commercial APIs from videos or images (e.g., Peng, 2018), 2) using supervised machine learning models to extract feature attributes defined by researchers (e.g., Joo & Steinert-Threlkeld, 2018), 3) clustering features extracted from a supervised model pre-trained on a large image dataset and fine-tuned to the target dataset (i.e., transfer learning) (Peng, 2021), and 4) computationally analyzing visual modalities of images or videos, such as color or saturation (Peng & Jemmott, 2018). The present paper follows the last approach, which focuses on analyzing visual attributes of conspiracy and debunking videos. ...
... Communication studies employing computational techniques to analyze visual frames follow at least four approaches (Peng, 2021): 1) extracting attributes predetermined by open-source computer vision libraries and commercial APIs from videos or images (e.g., Peng, 2018), 2) using supervised machine learning models to extract feature attributes defined by researchers (e.g., Joo & Steinert-Threlkeld, 2018), 3) clustering features extracted from a supervised model pre-trained on a large image dataset and fine-tuned to the target dataset (i.e., transfer learning) (Peng, 2021), and 4) computationally analyzing visual modalities of images or videos, such as color or saturation (Peng & Jemmott, 2018). The present paper follows the last approach, which focuses on analyzing visual attributes of conspiracy and debunking videos. ...
Preprint
Full-text available
Recent years have witnessed an explosion of science conspiracy videos on the Internet, challenging science epistemology and public understanding of science. Scholars have started to examine the persuasion techniques used in conspiracy messages such as uncertainty and fear yet, little is understood about the visual narratives, especially how visual narratives differ in videos that debunk conspiracies versus those that propagate conspiracies. This paper addresses this gap in understanding visual framing in conspiracy videos through analyzing millions of frames from conspiracy and counter-conspiracy YouTube videos using computational methods. We found that conspiracy videos tended to use lower color variance and brightness, especially in thumbnails and earlier parts of the videos. This paper also demonstrates how researchers can integrate textual and visual features for identifying conspiracies on social media and discusses the implications of computational modeling for scholars interested in studying visual manipulation in the digital era.
... Thus, exploring the role of images in political communication campaigns has gained great importance. Previous research on politicians' social media activities have been largely limited to Facebook and Twitter, and they mostly examine textual instead of visual content (Peng, 2021). Although leaders are increasingly using Instagram for self-presentation and expression, there are relatively few studies on their self-framing and self-presentation behaviours on the platform (Lalancette and Raynauld, 2019;Muñoz and Towner, 2017;Peng, 2021). ...
... Previous research on politicians' social media activities have been largely limited to Facebook and Twitter, and they mostly examine textual instead of visual content (Peng, 2021). Although leaders are increasingly using Instagram for self-presentation and expression, there are relatively few studies on their self-framing and self-presentation behaviours on the platform (Lalancette and Raynauld, 2019;Muñoz and Towner, 2017;Peng, 2021). ...
... Caple, 2020). On the other hand, the effects of visual attributes are less understood (see Peng, 2021) and analysis of visual social media constitutes challenges on a large scale (Peng, 2021;Rodriguez and Dimitrova, 2011). Russmann (2020) categorizes the methodological challenges of studying political actors on Instagram in three parts: temporal context of the visual materials, multidimensionality of the visuals and ethical considerations. ...
Article
This study aims to explore how political leaders used Instagram to execute self-presentation strategies in mayoral elections, including the dominant use of personalized tactics. The article reports findings of a visual framing analysis of 2,776 images featuring 2019 Istanbul mayoral election candidates Ekrem İmamoğlu (the Republican People’s Party, CHP) and Binali Yıldırım (the Justice and Development Party, AKP). The case is unusual because the initial election, which had resulted in İmamoğlu’s victory, was cancelled and a re-run was subsequently held. After many events, İmamoğlu succeeded again, becoming the first opposition politician to take control of Istanbul from the ruling AKP. In this study, we adapt Grabe and Bucy’s (2009) quantitative visual framing analysis to examine Instagram posts, from candidacy announcements until the election re-run. The results show that both candidates used the Ideal Candidate frame, with occasional increases in the frequency of the application of the Populist Campaigner frame. Self-frames in different time periods during this election are discussed, as well the frames that voters engaged with most frequently.
... Furthermore, scholars also applied image recognition to infer political ideology from social media images (Xi et al., 2020). Also focusing on political representatives, Peng examines the popularity factors of their social media posts, finding that self-personalization strategies increase audience engagement (Peng, 2020). At last, recent studies also applied image recognition to examine media representations. ...
Preprint
Full-text available
Digitization led to an enormous increase in the availability of visual data. As images are an important aspect of human communication, decades of social science research have analysed images, yet in mostly manual fashion with limited scaling capacities. In this work, we outline how recent advances in computer vision enable automated image analysis, allowing researchers to further unlock the potential of digital behavioural data. We introduce the field of computational social science and conduct a literature review of early studies using image recognition. We also highlight important aspects to be considered, such as computational demands and biases of computer vision models. Furthermore, in a case study, we examine the online behaviour of U.S. Members of Congress during the early COVID-19 pandemic in 2020. In particular, we focus on sharing images showing face masks as they are a crucial aspect of health and safety measures during the pandemic. Using Instagram data and models for detecting face masks, we find that temporal dynamics and party affiliation play a substantial role in the likelihood of sharing images of people wearing face masks: images with masks are more often posted after the introduction of mask mandates and Democratic party members are more likely to share images with masks. In addition, we find somewhat weaker to no differences regarding the age and gender of politicians.
... The computer vision community in computer science has already begun to turn their attention to unsupervised image analysis and has proposed several image clustering methods Frey and Dueck 2007;Guerin et al. 2017). A few social scientific studies have also applied clustering methods to images, demonstrating the potential of unsupervised methods in discovering meaningful patterns in visual content, such as content categories in Instagram images Manikonda and De Choudhury 2017;Peng 2021) and types of gestures in videos of politicians (Kang et al. 2020a). Still, the social scientific community lacks a comprehensive guide on how to perform image clustering on social scientific visual data and the advantages and caveats of different approaches. ...
Article
Automated image analysis has received increasing attention in social scientific research, yet existing scholarship has mostly covered the application of supervised learning to classify images into predefined categories. This study focuses on the task of unsupervised image clustering, which aims to automatically discover categories from unlabelled image data. We first review the steps to perform image clustering and then focus on one key challenge in this task—finding intermediate representations of images. We present several methods of extracting intermediate image representations, including the bag-of-visual-words model, self-supervised learning, and transfer learning (in particular, feature extraction with pretrained models). We compare these methods using various visual datasets, including images related to protests in China from Weibo, images about climate change on Instagram, and profile images of the Russian Internet Research Agency on Twitter. In addition, we propose a systematic way to interpret and validate clustering solutions. Results show that transfer learning significantly outperforms the other methods. The dataset used in the pretrained model critically determines what categories the algorithms can discover.
... Scholars have also applied deep learning models, such as Convolutional Neural Networks (CNNs), to predict image aesthetics (Talebi & Milanfar, 2018). Researchers have already started to use deep learning models to recognize and classify content themes in visual media (Joo et al., 2019;Joo & Steinert-Threlkeld, 2018;Peng, 2020;Zhang & Peng, 2021). Similarly, researchers can use deep learning models to predict the aesthetic appeal of visual media (Talebi & Milanfar, 2018) or high-level aesthetic features, such as color harmony and the rule of thirds (Kong, Shen, Lin, Mech, & Fowlkes, 2016). ...
Preprint
Full-text available
Visual aesthetics are related to a broad range of communication and psychological outcomes, yet the tools of computational aesthetic analysis are not widely available in the community of social science scholars. This article addresses this gap and provides a tutorial for social scientists to measure a broad range of hand-crafted aesthetic attributes of visual media, such as colorfulness and visual complexity. It introduces Athec, a Python library developed for computational aesthetic analysis in social science research, which can be readily applied by future researchers. In addition, a case study applies Athec to compare the visual aesthetics of Instagram posts from the two candidates in the 2016 US presidential election, Hillary Clinton and Donald Trump, showing how amateurishness and authenticity are reflected in politicians' visual messages. With tools of computational aesthetic analysis, communication researchers can better understand the antecedents and outcomes of visual aesthetics beyond the content of visual media.
Article
On July 17, 2021, the CDU's chancellor candidate Armin Laschet was photographed laughing during a speech by the German Federal President in the flood-stricken city of Erftstadt. The photographic images caused an uproar and contributed to the CDU's defeat in the September 23 election. The paper analyzes why these images resonated with such damaging effects. Theoretically, it sets the analysis on the background of the moralization and personalization of politics and argues that photography, with its ability to capture behavior at a distance, plays a prominent role in these processes. While this condition explains why an image of a laughing politician can generate such indignation in the first place, the paper discusses how this effect was amplified in the case of Laschet by a range of contextual features: (a) the timing of the images in the middle of an election period where politicians come under intense scrutiny; (b) their appearance in a crisis situation (the German flooding disaster) where politicians are surrounded by other role expectations than in routine periods; (3) Laschet's new, insecure position as leader of the CDU; (d) his history of scandals and poor political judgment; and (e) the frivolous and boisterous manner of his laughter. At a general theoretical level, the paper's insights caution us to avoid prima facie conclusions about the autonomous power of photographs. Instead, they encourage analytical sensitivity to the importance of timing and context as explanatory elements in our understanding of photographic exposés.
Article
This study demonstrates how localization and homogenization can co-occur in different aspects of smartphone usage. Smartphones afford individualization of media behavior: users can begin, end, or switch between countless tasks anytime, but this individualization is shaped by shared environments such that smartphone usage may be similar among those who share such environments but contain differences, or localization, across environments or regions. Yet for all users, smartphone screen interactions are bounded and guided by nearly identical smartphone interfaces, suggesting that smartphone usage may be similar or homogenized across all individuals regardless of environment. We study homogenization and localization by comparing the temporal, visual, and experiential composition of screen activity among individuals in three dissimilar media environments—the United States, China, and Myanmar—using one week of screenshot data captured passively every 5 s by the novel Screenomics framework. We find that overall usage levels are consistently dissimilar across media environments, while metrics that depend more on moment-level decisions and user-interface design do not vary significantly across media environments. These results suggest that quantitative research on homogenization and localization should analyze behavior driven by user interfaces and by contextually determined parameters, respectively.
Article
Current researchers pay less attention to the image position and layout of tweets containing multiple images. This study explored the impact of image position and layout on user engagement on the Weibo platform. The XGBoost model trained on single‐image tweet data was used to predict the “user engagement potential” of images in multi‐image tweets. Then, the image position and layout effects on user engagement were analyzed through correlation analysis and OLS regression. It was found that the right position was more important in tweets with less than or equal to 4 images, and the position effects became symmetric with image adding. Layouts with 2, 3, 4, 5, 6, 8 images had positive effects on user engagement, while layouts with 7 and 9 or more images had negative effects. This study provides insights for user engagement with social media images and may help improve interaction.
Article
Full-text available
This article is based on a content analysis of the 17,811 Instagram posts made by all 534 members of the United States Congress who were seated for the duration of the first 6 months of the 115th session. I find that women are significantly more likely than men to have an Instagram account. Senators and women post significantly more times to their accounts. And a member’s personal characteristics, such as their chamber, party, and age, had significant effects on the type of content posted to Instagram. I conclude that members of Congress use Instagram similarly to how they use other social media platforms, that parties in and out of power use Instagram in substantively different ways, and that the more personal accounts of younger members suggest future changes in Congressional representation.
Article
Full-text available
How do today’s partisan media outlets produce ideological bias in their visual coverage of political candidates? Applying computer vision techniques, this study examined 13,026 images from 15 news websites about the two candidates in the 2016 U.S. presidential election. The analysis unveils a set of visual attributes (e.g., facial expressions, face size, skin condition) that were adopted by media outlets of varying ideologies to differentially portray these two candidates. In addition, this study recruited 596 crowdsourced workers to rate a subset of 1,200 images and demonstrated that some visual features also effectively shape viewers’ perceptions of media slant and impressions of the candidates. For example, Clinton was portrayed with more expressions of happiness, which rendered her as more favorable, whereas Trump was associated with more expressions of anger, which made him look less positive but more dominant. These differences in facial expressions varied in line with media outlets’ political leanings.
Article
Full-text available
Do images affect online political mobilization? If so, how? These questions are of fundamental importance to scholars of social movements, contentious politics, and political behavior generally. However, little prior work has systematically addressed the role of images in mobilizing online participation in social movements. We first confirm that images have a positive mobilizing effect in the context of online protest activity. We then argue that images are mobilizing because they trigger stronger emotional reactions than text. Building on existing political psychology models, we theorize that images evoking enthusiasm, anger, and fear should be particularly mobilizing, while sadness should be demobilizing. We test the argument through a study of Twitter activity related to a Black Lives Matter protest. We find that both images in general and some of the proposed emotional attributes (enthusiasm and fear) contribute to online participation. The results hold when controlling for alternative theoretical mechanisms for why images should be mobilizing, and for the presence of frequent image features. Our paper provides evidence supporting the broad argument that images increase the likelihood of a protest to spread online while teasing out the mechanisms at play in a new media environment.
Article
Full-text available
The widely circulated food photos online have become an important part of our visual culture. Combining human ratings of food characteristics and computational analysis of visual aesthetics, we examined what contributed to the aesthetic appeal of a diversity of food photographs (N = 300) and likes and comments they received in an artificial newsfeed from participants (N = 399). The results revealed that people tended to like and share images containing tasty foods. Both healthy and unhealthy foods were able to gain likes. Aesthetic appeal and specific visual features, such as the use of arousing colors and different components of visual complexity, also influenced the popularity of food images. This work demonstrates the potential of applying computer vision methods in visual analysis, offers insights into image virality, and provides practical guidelines for communicating healthy eating.
Article
Full-text available
The present study aims to contribute to the agenda setting theory and political campaign literature by examining candidates’ tweets and their effects on voter reactions in the context of the 2016 U.S. presidential election. Content analysis of Donald Trump’s and Hillary Clinton’s 3-month tweets (N = 1575) revealed that half of their tweets were attacks, and those attacks were effective in attracting favorites and retweets for both candidates. Their tweets reflected their issue agendas highlighted on campaign websites, and they mainly emphasized issues owned by their parties in both venues. Some of the issues Trump stressed in his tweets (i.e., media bias and Clinton’s alleged dishonesty) drew significantly more favorites and retweets, suggesting public agenda setting possibilities through Twitter. None of the issues Clinton emphasized were significant predictors of favorites and retweets. However, visual elements such as pictures and videos were effective in bringing voter reactions for Clinton. While Clinton sent twice as many tweets as Trump did during the three months, Trump’s tweet received in average three times as many favorites and retweets as Clinton’s. Overall, the results show that Trump was more successful than Clinton in drawing public attention to preferred issues through Twitter.
Article
Full-text available
While research has gauged the degree to which political actors focus on their personal rather their more public sides in their communication efforts, few studies have assessed the extent to which personalized content succeeds in gaining traction among online followers. The current study does just that, focusing on the Instagram accounts operated by Norwegian parties and party leaders. Results indicate that party leaders emerge as more successful than parties in gaining attention through 'likes' and comments, and that they offer personalized content to higher degrees than the parties they represent. While personalized content might lead to increased political engagement among citizens, the fact that personalization 'works' in terms of gaining attention might also skew political PR and marketing towards excessive use of such themes.
Book
Images play a crucial role in shaping and reflecting political life. Digitization has vastly increased the presence of such images in daily life, creating valuable new research opportunities for social scientists. We show how recent innovations in computer vision methods can substantially lower the costs of using images as data. We introduce readers to the deep learning algorithms commonly used for object recognition, facial recognition, and visual sentiment analysis. We then provide guidance and specific instructions for scholars interested in using these methods in their own research.
Article
With the integration of social media in political communication repertoires, politicians now permanently campaign for support online. By promoting their personal agenda, politicians increasingly profile themselves independent from their associated parties on the web (i.e., self-personalization). By focusing on self-personalization as a multi-layered concept (i.e., professional, emotional, private self-personalization), this study investigates both the use and consequences of self-personalization on Facebook. A manual content analysis of politicians’ Facebook posts (N = 435) reveals that self-personalization is indeed often used as a communication style on Facebook and is most often present in visual communication. Moreover, the study shows that the use of a more emotional and private style provides a beneficial tool for politicians’ impression management. Publishing emotional and private content yields positive effects on audience engagement, suggesting audiences’ demand for more intimate and emotional impressions of public figures on the web.
Article
Political campaigns have been systematically using social media for strategic advantage. However, little is known about how competitiveness affects the ways candidates communicate online. Our study analyzes how race competitiveness as measured by polling performance influences candidates’ strategies on Twitter and Facebook. We analyze all social media messages of Republican and Democratic candidates in states that held gubernatorial elections in 2014 using supervised automated content analysis. We find that position in the polls and that race competitiveness are correlated with the ways candidates communicate on social media, and that candidates use Twitter and Facebook in different ways to communicate with the public.