See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/266655817

Faces engage us: photos with faces attract more likes and comments on
Instagram

Article · April 2014

DOI: 10.1145/2556288.2557403

CITATIONS READS
261 21,687

3 authors, including:

David Ayman Shamma

Toyota Research Institute
134 PUBLICATIONS 8,091 CITATIONS

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

CozyMaps View project

All content following this page was uploaded by David Ayman Shamma on 03 June 2016.

The user has requested enhancement of the downloaded file.

Session: Personal Values and Preferences CHI 2014, One of a CHInd, Toronto, ON, Canada

Faces Engage Us: Photos with Faces Attract

More Likes and Comments on Instagram
Saeideh Bakhshi David A. Shamma Eric Gilbert
Georgia Tech Yahoo Labs Georgia Tech
[email protected] [email protected] [email protected]

ACM Classification Keywords

H.5.m. Information interfaces and presentation (e.g., HCI):
ABSTRACT Miscellaneous.
Photos are becoming prominent means of communication
online. Despite photos’ pervasive presence in social media INTRODUCTION
and online world, we know little about how people interact Even as babies, humans love to look at faces; infants, barely
and engage with their content. Understanding how photo minutes old, turn toward faces, sensing that they are impor-
content might signify engagement, can impact both science tant [39]. It is widely accepted in neuroscience that face per-
and design, influencing production and distribution. One ception is perhaps the most highly developed human visual
common type of photo content that is shared on social media, skill [26]. Faces are also powerful channels of nonverbal com-
is the photos of people. From studies of offline behavior, we munication [45]. We constantly monitor faces because they
know that human faces are powerful channels of non-verbal provide vital clues in an impressive variety of contexts: attrac-
communication. In this paper, we study this behavioral tion, the complexity of emotions, identity, age, humor, and a
phenomena online. We ask how presence of a face, it’s age person’s regional and national background [24].
and gender might impact social engagement on the photo.
We use a corpus of 1 million Instagram images and organize Many of the faces we see everyday now have an online
our study around two social engagement feedback factors, presence. Photo sharing communities such as Instagram
likes and comments. Our results show that photos with faces have made it possible to communicate with large groups of
are 38% more likely to receive likes and 32% more likely to distributed people through an image—be it a picture of whats
receive comments, even after controlling for social network for dinner or a selfie—perhaps more easily than through
reach and activity. We find, however, that the number of words alone. As Kelsey puts it, “we are moving away from
faces, their age and gender do not have an effect. This photography as a way of recording and storing the past, and
work presents the first results on how photos with human instead turning photography into a social medium in its own
faces relate to engagement on large scale image sharing right” [35].
communities. In addition to contributing to the research Online photo sharing communities have grown at an impres-
around online user behavior, our findings offer a new line of sive pace. At the time of this writing, Instagram users up-
future work using visual analysis. load 55 million photos a day to the site1 . This presents a
key research challenge for photo sharing communities like
Author Keywords Instagram (cf. Flickr, Imgur, Tumblr): how do we discover
mobile; photo; Instagram; faces; face detection; engagement; the mechanisms by which users communicate around visual
age; demographics; gender; content; image; social media; content and engage with such content. In other words, since
image sharing community engagement is vital to photo sharing communities, it is crit-
ical to understand what form of content drives engagement.
This material is based upon work supported in part by the Defense Ad-
vanced Research Projects Agency (DARPA) under Contract No. W911NF- While several research studies have focused on how users en-
12-1-0043. Any opinions, findings and conclusions or recommendations ex- gage with textual content [6, 9, 11, 12, 31, 38], there are few
pressed in this material are those of the authors and do not necessarily reflect studies on what makes visual content socially engaging on-
the views of DARPA or the U.S. Government. The views and conclusions line. To investigate this, we ask the following research ques-
contained in this document are those of the authors and should not be inter-
preted as representing the official policies, either expressly or implied, of the
tions in this paper, driven by social psychology work on face
Defense Advanced Research Projects Agency or the U.S. Government. perception:
RQ1: Do photos with faces differ in online engagement com-
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
pared to photos without them?
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. Copyrights for components of this work owned by others than ACM
must be honored. Abstracting with credit is permitted. To copy otherwise, or republish,
RQ2: If so, how do characteristics of the image subject, such
to post on servers or to redistribute to lists, requires prior specific permission and/or a as gender and age, affect engagement?
fee. Request permissions from [email protected].
CHI 2014, April 26May 1, 2014, Toronto, Ontario, Canada. 1
Copyright 2014 ACM 978-1-4503-2473-1/14/04..$15.00. http://instagram.com/press (Accessed 9/2013)
http://dx.doi.org/10.1145/2556288.2557403

965
Session: Personal Values and Preferences CHI 2014, One of a CHInd, Toronto, ON, Canada

We use Instagram to answer our research questions. Insta- engagement. Explicit requests, personal testimonials relating
gram has over 150 million active monthly users who collec- one’s connection to the group, and staying on-topic increased
tively generate 1.2 billion likes per day. There are two social the probability of receiving a reply. Newcomers to a
aspects of engagement here: the number of likes and the num- group were less likely to receive a reply than veterans [6].
ber of comments on the Instagram image 2 . Using a corpus of Following up on this work, Burke et al. studied the role of
1 million images from the community, we find that on an av- self-disclosing introductions, mentions of the poster’s age,
erage a photo that contains a face receives 38% more likes and and an acknowledgment that this is the poster’s first post;
32% more comments compared to a photo that does not con- these factors were found to increase reply probability [11].
tain any faces (even after controlling for user activity levels In another study, Burke and Kraut, studied the effect of the
and social network reach). Further, we find that the number politeness of a post, finding that politeness leads to more
of faces in the photo, their age and their gender do not impact replies in certain types of groups, while in other types of
engagement. groups, rudeness actually increases replies [12]. On Twitter,
researchers have used retweeting as a measure of community
To our knowledge, our study is one of the first to show, sys- interest/engagement, and have investigated the features that
tematically and at scale, how photos with faces drive online predict retweeting. Suh et al. found that the presence of
social engagement. In addition to contributing to the ongoing URLs and hashtags in tweets predicted more retweeting, as
research conversation surrounding engagement, we believe did a richer connection with the community [44]. In a recent
that these findings create a path for future work, not only to work, Gilbert et. al. studied Pinterest- a social networking
uncover the impact of faces on other aspects of online user be- site based on images- and found that which properties of an
havior, but also to use computer vision techniques to discover image makes the content more interesting to users [23]. The
other antecedents of engagement. For example, we may be properties used in this work are based on meta data and not
able to apply vision techniques to relate facial expressions of
the content of images.
emotion to social behavior.
As far as we know, however, we have no such similar line
We begin with a review of related research on content and of work on how image content can affect different aspects of
face perception, and a summary of the Instagram community. online behavior, such as engagement, diffusion or link forma-
Next, we introduce the corpus we collected from Instagram tion. In our work, we intend to provide an understanding of
and describe the statistical methods we used to isolate the ef- image engagement by looking at the photo content.
fect of faces on likes and comments. Finally, we interpret our
findings within the frame of existing work, both in theory and
in practice. Face perception
One of the common types of photo content shared on social
LITERATURE REVIEW networking sites is the photos of people or the ones with
In this section, we describe related work on role of content human faces in them. Through daily experience, we know
in engagement, face perception theories and HCI studies of that human faces are readily distinguishable. People tend to
faces. We then summarize the media community of our study, find faces in unexpected scenes and photographs even where
Instagram. faces do not exist. For example, the 1976 Viking 1 prob pho-
tographed a shadowed region on Mars’ northern planes that
Role of Content on User Engagement
resembled a face. While higher resolution imagery has shown
the region to actually be a mesa, the face on Mars remains
As Ellison and colleagues note, “the primary function of these
a pop icon and the source of many books, TV shows, and
[social network] sites is to consume and distribute personal
films [40].
content about the self” [22]. Sharing content can in turn en-
sure that users remain engaged and committed in the future Faces have long been a source of scientific interest in a wide
visits [13, 47]. On the other hand, users have diverse motiva- range of disciplines. In recent years, this breadth of interests,
tions to share content on social network sites. For example, approaches and expertise has led directly to rapid advances in
users may share useful content to appear knowledgeable or our understanding of many different aspects of how we per-
simply to help out [50]. Not only the content of posts, but also ceive and process faces [10]. The human brain has evolved
the emotional valence behind it can drive its usage. For exam- to recognize faces within hours after birth [33]. Human in-
ple, in a recent study, researchers used New York Times arti- fants only minutes old attend particularly to face-like stimuli
cles to examine the relationship between the emotion evoked relative to equally complicated non-face stimuli [10, 30]. We
by content and its virality [9], finding that that there is a direct prefer to look at faces from that early age and thereafter, of-
relationship. ten opting to spend more time looking at faces than any other
type of object [53]. By the age of two months, infants begin
Much research attention has gone into investigating what
to differentiate specific visual features of the face [39] and
makes content in an online community interesting to its
process facial expressions [20]. Our brains have a specific re-
members. In a series of studies conducted on Usenet news-
gion, Fusiform Face Area (FFA), that is specialized for facial
groups, researchers investigated properties that influenced
recognition [34, 41]. Faces are important for social cognition
the likelihood of reply, a measure of the community’s
as well, not only because we are able to recognize them earlier
2
We use the terms image and photo interchangeably throughout this than other objects, but also because they display our feelings
paper to refer to the images on Instagram. about past, current and future events through expressions [17,

966
Session: Personal Values and Preferences CHI 2014, One of a CHInd, Toronto, ON, Canada

21]. This can be highly important to very practical concerns: Instagram is a social network site designed around photo-
faces, particularly attractive ones, are found to be effective in and video-sharing. It enables users to take photos and videos
improving consumer responses to advertisements [7]. with their mobile devices, apply digital filters to them and
share them on variety of social networking services, such as
Our research examines the presence of this phenomena on- Facebook, Twitter, Tumblr and Flickr [2], all of which are
line, by analyzing the effect of having faces in engaging users social media sites in their own right. Instagram has rapidly
on Instagram. gained popularity with over 100 million active users as of
April 2012 [19]. The total number of photographs uploaded
Faces and HCI research
recently exceeded one billion [1, 3].
In HCI research, there is a great deal of work exploring the
benefits of using face icons and faces in interfaces [36, 43, Instagram accounts are public by default, unless users elect
45, 49]. Walker et al. [49] studied how having faces and facial to create a private account; there is no tier privacy photo by
expressions for a computer application affects users’ perfor- photo. To add a photo, users can take a photo from inside the
mance and productivity. They compared subjects’ responses app. It is also possible to choose a photo from an existing al-
to an interview survey under three conditions: questions spo- bum on the mobile device to share with Instagram followers.
ken by a synthesized face with neutral expressions, spoken by Instagram users can apply filters on their photos. An Insta-
a face with stern expressions, or text only. Subjects who re- gram filter is a digital layer that when added to a photo, gives
sponded to the spoken face made more effort to answer the it the appearance of stylistic editing. Some filters enhance the
questions by spending more time, writing more comments colors in a photo, while others dull the light to a soft glow for
and making fewer mistakes. They reported that having a face an aged, vintage appearance.
is engaging and takes more effort and attention from the user.
Despite the popularity of Instagram, there is little scholarly
Takeuchi et al. [46] compared users’ impressions of an agent work on it. In a recent piece, Hochman et al. [29] analyzed
which helped them to win a card game. The agent was repre- colors in photos uploaded from two different cities of New
sented either as an arrow or a face. They showed that users re- York and Tokyo and found differences across the two loca-
spond differently to systems having a face than to those with- tions. For instance, hues of pictures in New York were mostly
out. The arrow was recognized as useful and reliable, while blue-gray, while those in Tokyo were characterized by domi-
the face was rated as fun and entertaining. They conclude that nant red-yellow tones.
a face in an interface captures more attention and people try
to interpret the meaning behind the expression. METHODS
We take a quantitative approach in this paper to investigate the
Studies on embodied interfaces showed similar results. relationship between faces and engagement. While engage-
Agents are visual digital representations of a computer ment can be quantified in various ways, we use two essential
interface often in the form of human-like faces [15]. In a aspects of content on Instagram that can signal for engage-
review study of embodied agents [18], authors reported ment: likes and comments. The number of likes signals for
that adding an embodied agent to an interface made the the extent to which the content is interesting to users and the
experience more engaging. number of comments quantifies the level of discussion on the
social network. In this section, we describe the data we col-
Role of Age and Gender lected from Instagram and how we detected faces and their
Age and gender have been studied extensively as factors af- age and gender; followed by clarifying our statistical methods
fecting social media use [8, 16, 27, 28]. Recent data3 shows and analysis process.
that women form a majority of Facebook and Twitter users, as
well as dominating Pinterest; however, men are the majority Face detection
of users on Google+ and LinkedIn. In a recent study Gilbert Face detection and recognition from images or video is a pop-
et. al. [23] found that females are more likely to receive re- ular topic in vision research and it has received lots of atten-
pins and fewer followers than males on Pinterest. Moreover, tion [48, 51]. A general statement of the problem of machine
Pew Internet Research [4] ran a survey to give marketers a recognition of faces is usually formulated as follows: given
clearer picture of who they can expect to reach on Instagram. still or video images of a scene, identify or verify one or more
According to the source, 28% of U.S. internet users aged 18 persons in the scene using a stored database of faces or facial
to 29 snapped and posted photos on the network in December features. The solution to the problem involves segmentation
2012. 14% of those aged 30 to 49 did the same, and very few of faces (face detection) from cluttered scenes and extraction
users older than 50 participated in any way on Instagram. of features from the face.
Inspired by previous research on disparities in internet usage While the current state of the art in face detection and recog-
and social network audience, we used age and gender vari- nition is highly accurate [32], we did not have access to an
ables to investigate whether they affect the number of likes implementation that can work for large scale image analysis.
and comments on photos. We therefore used a publicly available face detection API de-
veloped by Face++ [5]. We only use the detection modules,
Instagram as the goal of this paper is to find relationship between exis-
3
http://mashable.com/2012/07/04/ tence of faces and the social engagement. Face++ provides a
men-women-social-media/ (Accessed 9/2013) set of compact, powerful, and cross-platform vision services,

967
Session: Personal Values and Preferences CHI 2014, One of a CHInd, Toronto, ON, Canada

Figure 1. Example Face++ face detection and how we construct our variables. The photo used in this example is a photo under Creative Commons
license from Flickr.

which enabled us to use cutting-edge vision techniques. The Our goal in this paper is to investigate the role of photos in
API does not provide us with an estimation of accuracy, so predicting user engagement on Instagram. We chose num-
we turn to a crowd-sourced validation method to confirm the ber of likes and number of comments as two features that
accuracy of our face detector described later in the validation represent fundamental aspects of engagement on the site. An
section. overview of each of these variables is provided in Table 1.
Face++ provides us with an API that accepts the URL of Likes: Number of likes is a measure of engagement for the
an Instagram image and returns information about detected photo. It quantifies the number of distinct users who liked
faces. This information includes the position of the face in the photo. Like is a strong social signal on Instagram that
the image, as well as the detected gender and age range of all demonstrates the extent to which users liked the photo.
faces. We then reduce the dimensionality of data by convert-
ing the results into a binary space, where we mark only when Comments: Number of comments is another measure of en-
there is a face in an image. We also identify whether any of the gagement, or as Yew and Shamma [52] note, a measure of
faces in the image belong to certain age ranges. The three age explicit action on the content. The number of comments
ranges we consider in this paper are (1) children and teens- is the number of distinct comments posted on the photo.
younger than 18, (2) young adults- faces with age between The number of comments determines the extent to which
18 and 35, and (3) older adults- older than 35. To evaluate users discussed the photo and hence it can be considered as
the role of gender, we construct another binary feature which measure of discussion.
determines whether at least one female or one male face is in Predictor variables
the image. Figure 1 shows an example Face++ detection and In this paper, we use two major control variables to adjust for
how we construct our variables. the impact of social network reach and a user’s activity.

Data Control: user’s followers count. An Instagram photo is

Our goal is to obtain a random sample of photos from In- posted by an Instagram user. The nature of relationship
stagram. Even though Instagram provides us with a publicly on Instagram is follower/following. Users form a social
available API, gathering a random subset of photos is a chal- network based on “follow” relationships. When A follows
lenging task. We can either search for photos by location or B, B’s photos will show up in A’s photo-stream. The
query on the list of most recent popular photos. We opted to number of followers signals the social network reach. The
start with a set of 2,000 popular Instagram photos, collected more number of followers, the more people can see the
on November 2012. We then used snowball sampling [25] to photo and there is presumably a higher chance of receiving
collect the users and their followers as well as a random set likes and comments.
of their photos. Our dataset consists of 23 million Instagram Control: user’s photo count. Photo count is the feature we
photos and over 3 million Instagram users. To soften biases use to quantify a user’s activity on the site. It represents the
due to snowball sampling, we randomly selected 1.1 million number of photos on a user’s profile. The larger values of
photos from this data set. The snowball sampling method was photo count show the user has shared more content on the
necessary because Instagram does not provide any mecha- site; in other words the user is more active.
nism by which to monitor the global stream of photos. Fig-
ure 2 shows a detailed flowchart of data collection, evaluation As we discussed in related work section, faces are found to be
and analysis processes. effective stimuli [10, 30] in attracting people’s attention. We
use a binary variable as our predictor to account for presence
Response variables (dependent measures) of a face in the photo.

968
Session: Personal Values and Preferences CHI 2014, One of a CHInd, Toronto, ON, Canada

❞❛t❛ ✥☎✁✁❡✥t✄☎✆ ❡ ❛✁✂❛t✄☎✆ ❛✆❛✁✢s✄s

❯✌✐✞✓ ✧✞✌✕✝✓♦✝✡ ★✩✧✑ ❘✝✞✟✠✡☛☞ ✌✝✡✍☛✎✟ ✏✑✒✒✒ ❋✠♦ ✝☛☛ ✣▼ ✐✡✝✓✎✌✑ ✚♦✎✝✕✎✟ ✤✐✞✝♦☞
✚✠☛☛✎✚✕✎✟ ✏✑✒✒✒ ✍✠✍✘☛✝♦ ✐✡✝✓✎✌✙ ✐✡✝✓✎✌ ✔✠♦ ✡✎✕✖✠✟ ✎✗✝☛✘✝✕✐✠✞✙ ✔✎✝✕✘♦✎✌ ✔✠♦ ✔✝✚✎✑ ✝✓✎✑ ✓✎✞✟✎♦✙

❱✐✝ ✌✞✠✇✤✝☛☛ ✌✝✡✍☛✐✞✓✑ ❋✠♦ ✎✝✚✖ ✐✡✝✓✎✑ ✱ ❚✘✉✎♦✌ ✛✘✟✓✎✟ ◆✎✓✝✕✐✗✎ ✤✐✞✠✡✐✝☛ ✡✠✟✎☛
✚✠☛☛✎✚✕✎✟ ✏✪▼ ✡✠♦✎ ✐✡✝✓✎✌✙ ✝☛✓✠♦✐✕✖✡Õ✌ ✔✝✚✎ ✜ ✝✓✎ ♦✎✌✘☛✕✌✙ ✝✌✌✎✌✌✎✟ ✎✦✎✚✕ ✠✞ ✎✞✓✝✓✎✡✎✞✕✙

❘✝✞✟✠✡☛☞ ✌✝✡✍☛✎✟ ✕✠ ✣▼✙ ❋✝✚✎✑ ✝✓✎✑ ✓✎✞✟✎♦ ♦✎✌✘☛✕✌ ✗✝☛✐✟✝✕✎✟✙

❊✝✚✖ ✐✡✝✓✎ ✝✘✓✡✎✞✕✎✟ ✇✐✕✖

✍♦✠✫☛✎ ✟✝✕✝ ✜ ✝✞✝☛☞✬✎✟ ✔✠♦ ✔✝✚✎✌✙

Figure 2. An overview of the steps taken in this paper for collection, evaluation and analysis of the data.

Has face. For each Instagram photo, we determine whether The distribution and short summary of each of these features
at least one human face exists in the photo. This is a binary is provided in Table 1.
feature; when it is set to 1, there is at least one face in the
image, otherwise it is set to 0.
Other than presence of faces, we consider variables identify- Statistical methods
ing age and gender of them. Our age and gender variables are Next, we present statistical methods we used to model our
derived using face detection method. two dependent variables, number of likes and number of
comments. Both dependent variables are count variables. We
Has children and teens- has face < 18 years old. We use a model the number of likes and the number of comments using
binary feature to determine whether the photo has any faces negative binomial regression, on two classes of independent
in the age group <18 years old. The variable is set to 1 variables: the control variables (followers count and photos
when at least one of the identified faces in the image ap- count) and our variables of interest (related to existence of a
pears to be younger than 18 years old, and set to 0 other- face, age group of the face and gender of the face). Negative
wise. binomial regression is well-suited for over-dispersed
distributions of count dependent variable [14]. We use
Has young adults- has face > 18 and < 35 years old. This negative binomial regression instead of Poisson regression
is another age feature that is set to 1 when at least one of since the variance of the dependent variable is larger than the
the identifies faces in the image appears to be between 18 mean for both likes and comments. We use over-dispersion
and 35 years old, and it is set to 0 otherwise. to test whether Poisson or negative binomial regression
should be used. This test was suggested by Cameron and
Has older adults- Has face > 35 years old. Our final age Trivedi [14], and involves a simple least-squares regression
feature is to identify presence of older adults in the image. to test the statistical significance of the over-dispersion
If at least one of the faces in the image appears to be older coefficient.
than 35 years old, this variable is set to 1, 0 otherwise.
The regression coefficients β allow us to understand the ef-
Has female face. This feature is a binary feature reflecting fect of an independent variable on the number of likes and
whether there is a female face in the photo. When the vari- comments (note that to be able to compare coefficients, we z-
able is set to 1, the image has at least one female face, and score all numerical variables before performing regression).
it is set to 0 otherwise. For the variables with heavy tail distribution, such as follow-
ers count and photos count, we log transformed the variables
Has male face. This feature is a binary feature reflecting before performing regression. We use Chi-squared statistics
whether there is a male face in the photo. When the to find the statistical significance of our regression models,
variable is set to 1, the image has at least one male face, computing the reduction in deviance as compared to a null
and it is set to 0 otherwise. model.

969
Session: Personal Values and Preferences CHI 2014, One of a CHInd, Toronto, ON, Canada

Type Variable Description Distribution

likes∗ Number of likes on each photo. 0 50 100 150

Engagement
comments∗ Number of comments on each photo. 0.0 2.5 5.0 7.5 10.0

followers∗ Number of users who follow the photo’s 0 2500 5000 7500
Audience owner.
& Activity
photos∗ Number of photos shared by photo’s owner. 0 100 200 300 400 500

has face 1 if the photo contains a face, 0 otherwise. 0e+00 3e+05 6e+05 9e+05

has face < 18 years 1 if there is at least one face younger than 0e+00 3e+05 6e+05 9e+05
old 18, 0 otherwise.
has face ∈ [18, 35] 1 if there is at least one face with age be-
years old tween 0e+00 3e+05 6e+05 9e+05
Faces 18 and 35, 0 otherwise.

has face > 35 years old 1 if there is at least one face older than 35, 0e+00 3e+05 6e+05 9e+05
0 otherwise.

has female face 1 if there is at least one female face in the 0e+00 3e+05 6e+05 9e+05
photo, 0 otherwise.

has male face 1 if there is at least one male face in the 0e+00 3e+05 6e+05 9e+05
photo, 0 otherwise.
Table 1. Distributions of quantitative and binary variables used in this paper. Variables marked with ‘*’ are log-transformed. The red and blue lines
identify mean and median of the distribution, respectively. Orange refers to 1’s in the bar graphs. The engagement variables are our dependent measures.
Audience and activity variables are used as controls, and faces variables are the focal point of this study.

FACE DETECTION VALIDATION The results of the human validation are summarized in Ta-
As we mentioned in the previous section, we use Face++ API ble 2.
to detect faces in Instagram photos. Even though the currently
used face detection mechanisms are high in accuracy (over RESULTS
95%), we undertake an additional evaluation step to validate We use negative binomial regression to model the number of
and confirm the accuracy of our methods. For this purpose, likes and comments on photos. The results of the regression
we crowd-source a random sample of photos from our dataset are presented in Table 3. We use the Chi-squared Test to find
to Mechanical Turkers in order to verify the results of API. the significance of the regression model, by computing the
The validation process is as follows: we select a random sam- reduction in deviance from a null model. For our likes model,
ple of 2,000 images from our dataset. We then create tasks on we find reduction in deviance of χ2 = (5.2M − 1.2M ), or
Mechanical Turk, where the image is shown to five different 76%, on 8 degrees of freedom. The test rejected the null hy-
Turkers. We ask questions regarding the faces they see in the pothesis of a null model (p < 10−15 ); hence, the model is
image. They answer the questions about each image by iden- well-suited to characterize the effects of the described vari-
tifying how many human faces they see in the image, how ables.
many of them are female and how many are male. We then For our comments model, the reduction in deviance is χ2 =
ask Turkers to categorize the faces into different age groups (1.79M − 1.1M ), or 38%, on 8 degrees of freedom. The test
and report the number of people in each age group. We specif- rejected the null hypothesis of a null model (p < 10−15 ). The
ically ask Turkers to only report the human faces and avoid model for comments is also well-suited to characterize the
reporting the people they see in the picture if the faces are not effects of the independent variables.
visible.
We test coefficients of all independent variables for the null
We take the majority votes on each image and the results of hypothesis of a zero-valued coefficient (two-sided) and find
Mechanical Turkers evaluations are in agreement with API that the test rejects the null hypothesis (p < 10−5 ) in all cases.
output 97% ± 0.75% of the time. 93% ± 1.11% of the Turker
evaluations are in agreement in detecting faces with age range Effect of control variables
under 18, 96% ± 0.86% in age range agreement between 18 We use number of followers and number of photos as our con-
and 35, and 99% ± 0.44% in detecting ages over 35. Overall trol variables. As expected, the followers count has a large
the evaluation of our face detecting API shows high accuracy. positive effect on the number of likes and comments. This

970
Session: Personal Values and Preferences CHI 2014, One of a CHInd, Toronto, ON, Canada

Validation test Accuracy Margin of error Gender of the faces in the image does not show any strong
has face 97% 0.75% effect on the image’s engagement. Table 3 shows that the β
has female face 96% 0.86% coefficients for gender variables are of negligible size com-
has male face 96% 0.86% pared to some other features such as existence of a face.
has face < 18 years old 93% 1.11%
has face between 18 96% 0.86%
and 35 years old DISCUSSION
has face > 35 years old 99% 0.44% Using Instagram as our research context, we set out to investi-
gate how photos with faces relate to engagement, as measured
Table 2. Results of Mechanical Turk evaluation for our face detection
approach. Margin of Error is computed for 95% confidence. Our face
as the number of likes and comments, compared to those
detector works correctly 97% ± 0.75% of the time. without. We considered presence of a face in a photo, it’s gen-
der and age as predictors, controlling for social network reach
means the higher the number of followers, the more likely it and activity. From this we asked two research questions: are
is for the photo to receive likes and comments. The higher photos with faces more engaging than those without and if so
number of followers guarantees a larger audience and so the how do the characteristics of a face in a photo affect engage-
photo is expected to be seen by more number of people, in- ment?
creasing the likelihood of receiving likes and comments.
As expected, we find that among the factors we measured,
On the other hand the number of photos shared by user shows the number of followers is the main driver of engagement for
a negative effect on both likes and comments. The number of both likes and comments on the photo. The number of fol-
photos is an indicator of activity on Instagram. As we can see lowers is a proxy for the size of a user’s audience. Having a
in our results (Table 3) the higher activity (number of pho- larger audience increases the likelihood of a like or comment,
tos), the lower chances of receiving likes and comments. This a common sense fact realized in our models. Furthermore, we
might also be interpreted another way: the more photos a user find that activity level is negatively correlated with likes and
has, the lower probability any single one has of being liked or comments. The more photos a user posts, the less likely it is
commented on. that her photos receive likes and comments. As we mentioned
earlier, this most likely represents the intuition that the more
Effect of faces photos a user posts, the less likely any one of them is to be
All other predictors in our model come from the face highly liked or commented.
detection results. We are interested in quantifying the
effect of faces and their comparative importance on social Faces engage us
engagement. We use a binary variable that reflects the The major finding of this paper is that the existence of a face
existence of a face in the image. We can see in Table 3 in a photo significantly affects its social engagement. This
that number of likes and comments are significantly higher effect is substantial, increasing the chances of receiving likes
when there is at least one face in the image (βlikes = 0.32, by 38% and comments by 32%. We also find that number
βcomments = 0.28, p < 10−15 ). This means that photos with of faces in the image does not have significant impact on
faces are 38% more likely (IRR = 0.38) to receive likes and engagement. Having a photo with a face, regardless of how
32% more likely (IRR = 0.32) to receive comments4 ). many faces are in the photo, increases the likelihood of re-
We also check the effect of number of faces on engagement ceiving likes and comments. Our findings connect to the find-
and find that while existence of a face positively correlates ings from offline studies in psychology, marketing and social
with the number of likes and comments, the number of behavior, as well as qualitative studies of HCI, confirming
faces does not particularly change this effect. Regardless of that people engage more with photos of faces.
whether it is a group photo or a single person’s photo, the fact
that a face is in the image significantly impacts the number of Age and gender do not impact engagement
likes and comments. It does not matter how many faces are Our results show that the age and gender of faces in the photo
in the image. We did not include the number of faces in the does not seem to drive or hinder it’s engagement values. This
final model to avoid co-linearity of the predictor variables. is a surprising finding, given the bias of demographics using
the site and the general belief that photos of kids or female
Effects of age and gender faces may get more attention. For comments, we see in the re-
To test whether the demographic of users [4] biases toward sults that there is a small negative effect of older adult photos.
photos with younger face groups, we considered using three Since the comments are mostly related to the extent to which
binary variables each identifying the age of a face. Table 3 a photo is discussed, the lower number of comments on this
shows that the age group of the faces are generally not strong type of photos can be related to the lower demographics of
predictors for the number of likes. In case of number of com- older adults on Instagram. Future work can look at effect of
ments the photos with adult age groups negatively affects the similar factors on other photo sharing communities such as
number of comments. This could be related to lower presence Pinterest with biased gender demographics.
of older age groups in the social network of Instagram.
4 Implications and future work
We use IRR to refer to Incidence Rate Ratio. We compute IRR for a
categorical independent variables x as the ratio of amount of change This work raises many fundamental questions about the na-
in the dependent variable (outcome) for x relative to a reference level ture of social interaction around multimedia content. We be-
of x. lieve that this is an initial step and there is a rich landscape

971
Session: Personal Values and Preferences CHI 2014, One of a CHInd, Toronto, ON, Canada

Variable β Std.Err p Variable β Std.Err p

−15
number of followers 1.32 0.00 < 10 number of followers 0.97 0.00 < 10−15
number of photos -0.21 0.00 < 10−15 number of photos -0.12 0.00 < 10−15
has face 0.32 0.01 < 10−15 has face 0.28 0.00 < 10−15
has face <18 years old 0.02 0.01 < 10−15 has face <18 years old -0.01 0.01 < 10−15
has face >18 and <25 -0.03 0.01 < 10−15 has face >18 and <25 -0.07 0.01 < 10−15
has face >25 years old -0.03 0.01 < 10−15 has face >25 years old -0.04 0.01 < 10−15
has female face -0.04 0.01 < 10−8 has female face -0.01 0.01 < 10−4
has male face -0.02 0.01 < 10−3 has male face -0.02 0.01 < 10−6
(Intercept) 3.47 0.00 < 10−15 (Intercept) 3.47 0.00 < 10−15
Null deviance 5208940 Null deviance 1790136
Residual deviance 1227787 Residual deviance 1105145
Table 3. Results of negative binomial regression with number of likes (left) and number of comments (right) as dependent variable.

of research directions and open questions in this area. Fu- The context in which faces appear also invites interesting
ture work can look at other visual characteristics of multime- questions about individual and group behavior. Are photos
dia and study their impact on online behavior. Here we find of friends group more/less popular than the family ones?
that faces might have an impact on engagement, but faces are What about selfies and people’s reaction to self portraits? It
just one visual feature. Other signals can be gathered from is worth studying the cultural impacts on photo sharing, say
people in photos, including facial expression, gaze direction, for example are group photos more engaging in collectivism
as well as, body posture and movement. Although facial ex- cultures rather than individualistic ones?
pressions reliably signal the so-called basic emotions such as
fear or happiness, human viewers are also surprisingly adept Camera-phones and mobile photo capture has changed how
at making reliable judgments about social information from we perceive photo-work in the academic community. This
impoverished stimuli, such as faint changes in facial expres- work takes one of the first steps into understanding modern
sions [30]. Emotional expressions in faces are known to ac- photo capture and consumption through the study of Insta-
tivate several areas of the brain [24]. Future work can look gram. That said, Instagram is one online ecosystem and it has
been claimed that perception and semantics in social media
at emotional expressions of faces and explore the effects on
sites is a construction of the community on that site [42]. For
user behavior. For example, are we more likely to comment
example, Instagram is a people-centric site and the influence
on wry smiles or broad grins?
of faces might be different in a product-centric site such as
Our quantitative results illuminate what is the response to the Pinterest. On the other hand a community such as Instagram,
photos with faces, but not why users behave this way or what which is strongly based on social connections might react dif-
kind of connections they make with such photos. Additional ferently to faces than a professional photography community
work, particularly using qualitative methods, is needed to an- such as Flickr.
swer these questions. Some of the most compelling questions
The practical implication of social engagement in online
concern the person in the photo; for example, are users en-
gaging with faces as generic objects or are they connecting photo sharing lies strongly in search and recommendation.
with the face as a person they know. Knowing photos with faces increase engagement suggests
one could increase their search ranking to keep people on
As our work is based on quantitative studies and observa- site and active. Our results highlight the importance of
tional data, we cannot make any strong causal claims. We effective methods that take advantage of presence of faces
find that photos with faces have higher chances of being liked in photos for personalization of site content. Additionally,
and commented on, but we don’t know if faces are the exact while we have seen face finding applications for social media
cause of this. More experimental work needs to corroborate sites [37], these tools have been designed for the utility of
these findings. Further, the statistical methods we used exam- retrieval and not for conversation and comments.
ine only a small segment of behavior on the site.
For designers, the present findings may shed light on how
Faces and their presence connect to psychological studies of to filter, prioritize and highlight photos from the global im-
human behavior, and emphasize the importance of engaging age stream, especially ones that have just been submitted and
our unconscious perceptual biases—instantiated in this work therefore haven’t had time to accumulate very many likes and
as face perception. Future work can investigate the relation- comments.
ship between face perception theories and other aspects of
online user behavior. For example are faces effective when it CONCLUSION
comes to spreading the content on the social network? Are Faces are shown to be powerful visual tool used in human
photos or topics, accompanied by human faces more/less per- non verbal communication. With the widespread use of im-
suasive in terms of delivering the content? age sharing communities, most of which are on top of social

972
Session: Personal Values and Preferences CHI 2014, One of a CHInd, Toronto, ON, Canada

platforms, a key challenge in research community is to under- conference on Computer supported cooperative work,
stand the role of the image content in online user behavior. In ACM (2008), 281–284.
this paper, we took a first step toward uncovering an impor- 13. Burke, M., Marlow, C., and Lento, T. Feed me:
tant feature of some of images, the human faces. We find that motivating newcomer contribution in social network
photos with faces are 38% more likely to be liked and 32% sites. In Proceedings of the SIGCHI conference on
more likely to be commented on. Our results, however, show human factors in computing systems, acm (2009),
that number of faces, their age and gender do not have signif- 945–954.
icant impact. In addition to speaking to the ongoing studies
in online user behavior and social engagement, our findings 14. Cameron, C. A., and Trivedi, P. K. Regression analysis
open a new thread of future work, suggesting research in vi- of count data (econometric society monographs).
sual analysis. Cambridge university press, September 1998.
15. Cassell, J., et al. Nudge nudge wink wink: elements of
REFERENCES face-to-face conversation for embodied conversational
1. 150,000,000 photos. http://blog.instagram.com/ agents. Embodied Conversational Agents (2000), 1–27.
post/8758450298/150-000-000-photos, August 2011.
16. Cunha, E., Magno, G., Almeida, V., Gonçalves, M. A.,
2. Here’s how to use instagram. http://www. and Benevenuto, F. A gender based study of tagging
businessinsider.com/instagram-2010-11?op=1, May behavior in twitter. In Proceedings of the 23rd ACM
2011. conference on Hypertext and social media, ACM (2012),
3. Testing, testing... http://blog.instagram.com/post/ 323–324.
8758396471/testing-testing, July 2011.
17. Darwin, C. The expression of the emotions in man and
4. A demographic portrait of users of various social media animals. Oxford University Press, 1998.
services. http://pewinternet.org/Reports/2013/
18. Dehn, D. M., and Van Mulken, S. The impact of
Social-media-users/
animated interface agents: a review of empirical
Social-Networking-Site-Users/Demo-portrait.
research. International journal of human-computer
aspx, 2013.
studies 52, 1 (2000), 1–22.
5. Face++. http://en.faceplusplus.com/, 2013.
19. Desmarais, C. Facebook’s instagram says it has 90
6. Arguello, J., Butler, B. S., Joyce, E., Kraut, R., Ling, million monthly active users, February 2013.
K. S., Rosé, C., and Wang, X. Talk to me: foundations
20. Diener, E., Fraser, S. C., Beaman, A. L., and Kelem,
for successful individual-group interactions in online
R. T. Effects of deindividuation variables on stealing
communities. In Proceedings of the SIGCHI conference
among halloween trick-or-treaters. Journal of
on Human Factors in computing systems, ACM (2006),
personality and social psychology 33, 2 (1976), 178.
959–968.
21. Ekman, P., and Friesen, W. Pictures of facial affect.
7. Baker, M. J., and Churchill J., G. A. The impact of
consulting psychologists press, 1975.
physically attractive models on advertising evaluations.
Journal of marketing research (1977), 538–555. 22. Ellison, N., Steinfield, C., and Lampe, C. Connection
strategies: social capital implications of
8. Beckwith, L., Kissinger, C., Burnett, M., Wiedenbeck,
facebook-enabled communication practices. New media
S., Lawrance, J., Blackwell, A., and Cook, C. Tinkering
& society 13, 6 (2011), 873–892.
and gender in end-user programmers’ debugging. In
Proceedings of the SIGCHI conference on Human 23. Gilbert, E., Bakhshi, S., Chang, S., and Terveen, L. I
Factors in computing systems, ACM (2006), 231–240. need to try this?: a statistical overview of pinterest. In
9. Berger, J., and Milkman, K. L. What makes online Proceedings of the SIGCHI conference on human
content viral? Journal of marketing research 49, 2 factors in computing systems, ACM (2013), 2427–2436.
(2012), 192–205. 24. Goldman, A. I., and Sripada, C. S. Simulationist models
10. Bruce, V., and Young, A. In the eye of the beholder: the of face-based emotion recognition. Cognition 94, 3
science of face perception. Oxford University Press, (2005), 193–213.
1998. 25. Goodman, L. A. Snowball sampling. The Annals of
11. Burke, M., Joyce, E., Kim, T., Anand, V., and Kraut, R. Mathematical Statistics 32, 1 (1961), 148–170.
Introductions and requests: Rhetorical strategies that 26. Haxby, J., Hoffman, E., and Gobbini, M. The distributed
elicit response in online communities. In Communities human neural system for face perception. Trends in
and Technologies 2007. Springer, 2007, 21–39. cognitive sciences 4, 6 (2000), 223–233.
12. Burke, M., and Kraut, R. Mind your ps and qs: the 27. Herring, S. C. Gender and power in on-line
impact of politeness and rudeness in online communication. The handbook of language and gender
communities. In Proceedings of the 2008 ACM (2008), 202.

973
 Session: Personal Values and Preferences CHI 2014, One of a CHInd, Toronto, ON, Canada

28. Herring, S. C., Kouper, I., Scheidt, L. A., and Wright, 41. Sergent, J., Ohta, S., and Mdonald, B. Functional
E. L. Women and children last: The discursive neuroanatomy of face and object processing: a position
construction of weblogs. In Into the blogosphere: emission tomography study. Brain 115, 1 (1992), 15–36.
Rhetoric, community, and culture of weblogs, Citeseer
42. Shamma, D. A., Shaw, R., Shafton, P. L., and Liu, Y.
(2004).
Watch what i watch: using community activity to
29. Hochman, N., and Schwartz, R. Visualizing instagram: understand content. In MIR ’07: Proceedings of the
Tracing cultural visual rhythms. In Proceedings of the international workshop on Workshop on multimedia
workshop on social media visualization (socmedvis) in information retrieval (2007), 275–284.
conjunction with the sixth international AAAI conference
on weblogs and social media (ICWSM-12) (2012). 43. Sproull, L., Subramani, M., Kiesler, S., Walker, J. H.,
and Waters, K. When the interface is a face.
30. Iacoboni, M., Koski, L., Brass, M., Bekkering, H., Human-Computer Interaction 11, 2 (1996), 97–124.
Woods, R., Dubeau, M., Mazziotta, J., and Rizzolatti, G.
Reafferent copies of imitated actions in the right 44. Suh, B., Hong, L., Pirolli, P., and Chi, E. H. Want to be
superior temporal cortex. Proceedings of the national retweeted? large scale analytics on factors impacting
academy of sciences 98, 24 (2001), 13995–13999. retweet in twitter network. In 2010 IEEE Second
International Conference on Social Computing
31. Jamali, S., and Rangwala, H. Digging digg: comment (SocialCom) (2010), 177–184.
mining, popularity prediction, and social network
analysis. In International Conference on Web 45. Takeuchi, A., and Nagao, K. Communicative facial
Information Systems and Mining, 2009. WISM 2009, displays as a new conversational modality. In
ieee (2009), 32–38. Proceedings of the INTERSECT’93 and CHI’93
conference on human factors in computing systems,
32. Jenkins, R., and Burton, A. M. 100% accuracy in ACM (1993), 187–193.
automatic face recognition. Science 319, 5862 (2008),
435. 46. Takeuchi, a., and naito, t. Situated facial displays:
towards social interaction. In Proceedings of the SIGCHI
33. Johnson, M. H., Dziurawiec, S., Ellis, H., and Morton, J. conference on human factors in computing systems, acm
Newborns’ preferential tracking of face-like stimuli and press/addison-wesley publishing co. (1995), 450–455.
its subsequent decline. cognition 40, 1 (1991), 1–19.
47. Vasalou, A., Joinson, A., and Courvoisier, D. Cultural
34. Kanwisher, N., McDermott, J., and Chun, M. M. The differences, experience with social networks and the
fusiform face area: a module in human extrastriate nature of true commitment in facebook. International
cortex specialized for face perception. The Journal of journal of human-computer studies 68, 10 (2010),
Neuroscience 17, 11 (1997), 4302–4311. 719–728.
35. Kelsey, R. E., and Stimson, B. The meaning of 48. Viola, p., and jones, m. j. Robust real-time face
photography. Clark Art Institute, 2008. detection. International Journal of Computer Vision 57,
36. Laurel, B. Interface agents: Metaphors with character. 2 (2004), 137–154.
Human values and the design of computer technology
49. Walker, J. H., Sproull, L., and Subramani, R. Using a
(1997), 207–219.
human face in an interface. In Proceedings of the
37. Lei, Y.-H., Chen, Y.-Y., Iida, L., Chen, B.-C., Su, H.-H., SIGCHI conference on human factors in computing
and Hsu, W. H. Photo search by face positions and facial systems, ACM (1994), 85–91.
attributes on touch devices. In Proceedings of the 19th
ACM international conference on Multimedia, MM ’11, 50. Wojnicki, A., and Godes, D. Word-of-mouth as
ACM (New York, NY, USA, 2011), 651–654. self-enhancement. HBS marketing research paper,
06-01 (2008).
38. Millen, D. R., and Patterson, J. F. Stimulating social
engagement in a community network. In Proceedings of 51. Wright, J., Yang, A., Ganesh, A., Sastry, S., and Ma, Y.
the 2002 ACM conference on Computer supported Robust face recognition via sparse representation.
cooperative work, ACM (2002), 306–313. Pattern analysis and machine intelligence, IEEE
Transactions on 31, 2 (2009), 210–227.
39. Morton, J., and Johnson, M. H. Conspec and conlern: a
two-process theory of infant face recognition. 52. Yew, J., and Shamma, D. A. Know your data:
Psychological review 98, 2 (1991), 164. Understanding implicit usage versus explicit action in
video content classification. IS&T/SPIE Electronic
40. Phillips, T. Unmasking the face on mars. new Imaging (January 2011) (2011).
high-resolution images and 3d altimetry from nasa’s
mars global surveyor spacecraft reveal the face on mars 53. Young, A., Mcweeny, K., Hay, D., and Ellis, A.
for what it really is: a mesa. Matching familiar and unfamiliar faces on identity and
http://science.nasa.gov/science-news/ expression. Psychological research 48, 2 (1986), 63–68.
science-at-nasa/2001/ast24may_1/, May 2001.

974

View publication stats