RESEARCH ARTICLE

Cross-platform and cross-interaction study of

user personality based on images on Twitter
and Flickr
Zahra Riahi Samani1, Sharath Chandra Guntuku2,3,4*, Mohsen Ebrahimi Moghaddam1,
Daniel Preoţiuc-Pietro2,4, Lyle H. Ungar4
1 Faculty of Computer Science and Engineering, Shahid Beheshti University G.C, Tehran, Iran, 2 Positive
Psychology Center, University of Pennsylvania, Philadelphia, PA, United States of America, 3 School of
Medicine, University of Pennsylvania, Philadelphia, PA, United States of America, 4 Computer & Information
a1111111111 Science, University of Pennsylvania, Philadelphia, PA, United States of America
a1111111111
a1111111111 * [email protected]
a1111111111
a1111111111
Abstract
Assessing the predictive value of different social media platforms is important to understand
OPEN ACCESS
the variation in how users reveal themselves across multiple platforms. Most social media
platforms allow users to interact in multiple ways: by posting content to the platform, liking
Citation: Samani ZR, Guntuku SC, Moghaddam
ME, Preoţiuc-Pietro D, Ungar LH (2018) Cross- others’ posts, or building a user profile. While prior studies offer insights into how language
platform and cross-interaction study of user use differs across platforms, differences in image usage is less well understood. In this
personality based on images on Twitter and Flickr. study, we analyzed variation in image content with user personality across three interaction
PLoS ONE 13(7): e0198660. https://doi.org/
types (posts, likes and profile images) and two platforms, using a unique data set of users
10.1371/journal.pone.0198660
who are active on both Twitter and Flickr. Usage patterns on these two social media plat-
Editor: Kim-Kwang Raymond Choo, University of
forms revealed different aspects of users’ personality. Cross-platform data fusion is thus
Texas at San Antonio, UNITED STATES
shown to improve personality prediction performance.
Received: December 21, 2017

Accepted: May 23, 2018

Published: July 11, 2018

Copyright: © 2018 Samani et al. This is an open

access article distributed under the terms of the
Creative Commons Attribution License, which
permits unrestricted use, distribution, and
Introduction
reproduction in any medium, provided the original According to a Pew Research study [1], 56% of US adults online use more than one social
author and source are credited. media platform. While some of these, such as LinkedIn have a specific use [2], other platforms
Data Availability Statement: All result files are such as Twitter are used in diverse ways by different groups of users [3]. Also, there are multi-
available at DOI: 10.6084/m9.figshare.6469577. ple ways in which users can interact with a social media platform—either by posting content
The result files that we made available on the to the platform, liking content that others have posted or maintaining up their user profile.
repository contain the data necessary to reproduce
The volume and diversity of content that users produce and exchange on social media has
the Figures contained in the document. The
authors are not authorized, however, to share the
led to the possibility of performing computational analysis and prediction of users’ personality
individual-level Twitter data because it would with based on their social media footprints [4]. While several studies focused on one social media
be an Twitter TOS violation. platform and type of interaction, such as liked pages on Facebook [5], very few studies consid-
Funding: LHU acknowledges the support of the
ered cross-platform data to analyze personality differences [6, 7]. Moreover, no study to date
Templeton Religion Trust, grant TRT-0048. The examined the different types of interactions (termed as ‘modalities’ in the rest of this text) per-
funders had no role in study design, data collection formed on the same platform such as posted, liked and profile content.

PLOS ONE | https://doi.org/10.1371/journal.pone.0198660 July 11, 2018 1 / 19

 Predicting user personality on Twitter and Flickr from posted, liked and profile images

and analysis, decision to publish, or preparation of With images gaining popularity in social media posts, personality traits can be inferred
the manuscript. based on image-based content analysis. Images contain various concepts such as scenes,
Competing interests: Regarding Competing objects, colors or faces and these can be automatically captured using current computer vision
Interests, funding received from the Templeton algorithms. These representations can be used to analyze the relationship between users’ per-
Trust does not alter our adherence to PLOS ONE sonality and image posting across different modalities and social media platforms.
policies on sharing data and materials.
Prior research [8] suggests that personality is strongly expressed on a platform which offers
users sufficient self-expression and freedom of control. Social media platforms offer users the
opportunity to have multiple types of interactions. These modalities reveal more complex and
diverse patterns of behavior. Consequently, exploring different interactions that users have on
social media platforms might provide a better understanding of users’ personality.
The aim of this paper is to quantify image sharing preferences and to build models that
automatically predict users’ personality in a cross-/modal and cross-platform setting.
Research Questions: The research questions motivating this study are:
• How are personality traits related to what images users post, like and set as profile picture? We
term these as different modalities of interaction with the platform.
• How are personality traits expressed differently across platforms through images? We use a set
of users who have active accounts on both Flickr and Twitter.
• Can combining data from multiple platforms help improve performance of automatically pre-
dicting user personality?
Computational models that predict user traits based on their online footprints have several
applications in targeted online marketing, increasing acceptance of HCI systems, personalized
search and recommendations and exploring social science hypotheses based on large-scale
social media data.

Related work
With proliferation of mobile technologies and image sharing platforms, sharing pictures is the
most commonly action (82% of the American users), followed by exchanging text messages
(80% of the users) and accessing the Internet (56% of the users) [9]. In other words, “photos
have become an important social content online [10, 11] that and can serve as a substitute for
more direct forms of interaction like email [12].
This work contributes to recent social media trends that try to consider the content of their
users’ interaction to predict personality of their users. Personality is a combination of all the
attributes which includes differences in human behavior, thinking and feeling. Identifying per-
sonality of people has always been of great interest due to its importance. Personality traits
have influence on many aspects of user behavior such as job performance [13], music prefer-
ences [14], psychological conditions [15–17], leadership ability [18], academic abilities and
motivation [19], emotional responses to multimedia [20, 21], sales ability [13], perception of
multimedia quality [22–24] and so on.
Recent research has examined the interplay between users’ personality traits—usually mea-
sured using the Big Five model [25, 26]—and their social media data [27]. Facebook page likes
[5] and status updates [28] were used to accurately infer users’ personal information. Users
choice to disclose particular sections of their social media profile was used to study their per-
sonality traits [29]. Images on social media are now increasingly being used due to their
increased production and exchange in the recent years [4].
Most of previous research on images focused solely on profile images using facial features.
Self-assessed personalities of 100 users were predicted using their Facebook profile images [30]

PLOS ONE | https://doi.org/10.1371/journal.pone.0198660 July 11, 2018 2 / 19

 Predicting user personality on Twitter and Flickr from posted, liked and profile images

Table 1. Characteristics of recent work in Image-based personality analysis on social media.

Study # Users Assessment Image Type Social Platform Image Features
Ferwerda et al. 193 Perceived Posted Images Instagram Photos Content
[34] Personality
Ferwerda et al. 113 Self-assessed posted images Instagram Photos Colors, #Faces, Filters
[29]
Liu et al. [33] 66,502 Self-assessed and Profile images Twitter Color, Facial
+ 434 estimated
Nie et al. [35] 2238 Perceived Portrait Images Google Facial, Social information
Guntuku et al. 4132 Estimated and self- Posted, liked images Twitter Color, Bag of Imagga tags, VGG-Net trained on 1000 object
[36] + 161 assessed and text categories
Guntuku et al. 123 Self-assessed and Selfies (self-portraits) Weibo Color, Aesthetics, GIST, LBP, Bag-of-Visual-Words, Abstract
[37] perceived sentiments, Fisher encodings of SIFT, SURF, HOG and MSER
Guntuku et al. 300 Self-assessed and Liked (‘Fave’) images Flickr Colors, semantic features, aesthetics
[38, 39] perceived
Wei et al. [40] 3,162 Self-assessed Profile and posted Weibo Colors, CNN, text
images
Nie et al. [41] 1000 Social Media Portrait images Micro Blogs Concept and Emotion Detector, Active period, Level of attention
Behavior (interests), and frequency of posts and forwards
Sang et al. [42] 300 Self-assessed Liked (‘Fave’) images Flickr Aesthetics and Content feature
Segalin et al. [43] 300 Self-assessed Liked (‘Fave’) images Flickr Convolutional Neural Network
Segalin et al. [44] 300 Self-assessed Liked (‘Fave’) images Flickr Color, Composition, Texture, No of Faces
Segalin et al. [45] 11,736 Self-assessed Profile images Facebook Aesthetics, BOVW, VGG-Net, IATO
Skowron et al. [6] 62 Self-assessed Posted images, text Twitter and Pleasure-Arousal-Dominance, Color, Face, Body, Textual tags,
users’ meta features Instagram followers and publicly available counts
Xiong et al. [46] 300 Self-assessed Liked (‘Fave’) images Flickr Faces, Color, Composition, Texture, Gist
AlMoubayed 829 Perceived Face Images Face Recognition Eigenfaces
et al. [32] personality Database
Celli et al. [30] 112 Self-assessed Profile Images Facebook Bag of Visual Words (BOVW)
https://doi.org/10.1371/journal.pone.0198660.t001

with *65% accuracy using bag-of-visual-words features. Random portraits from the web [31]
and existing face recognition data sets [32] were also used to model users’ personality.
Recently, aesthetic features in addition to facial features were also used to study and predict
personality on a *66,000 user data set [33] from Twitter.
Multiple platforms have recently been studied to infer users’ personality. For instance, an
attempt to fuse cues from Instagram and Twitter reported a consistent decrease of the predic-
tion errors for each personality trait [6]. Also in a cross-platform scenario involving Instagram
and Twitter, [7] studied the differences in topics. However, no prior study examined the differ-
ent types of user interactions such as posts, likes and profile images.
Table 1 shows a comparison of most recent studies in this area. They use variety of image
features from different modalities of activity such as (profile, posted or liked) on different plat-
forms to predict personality. In this paper we do a cross-platform and cross modality analysis
to predict personality from social activities, comparing and contrasting the predictive value of
each.

Materials and methods

We use two data sets in our experiments. The first data set contains a set of Flickr users with
their self-assessed personality traits. This data set is used to compare the predictive power of

PLOS ONE | https://doi.org/10.1371/journal.pone.0198660 July 11, 2018 3 / 19

 Predicting user personality on Twitter and Flickr from posted, liked and profile images

Fig 1. Overview of cross-modal analysis.

https://doi.org/10.1371/journal.pone.0198660.g001

various image interactions of these users on Flickr. The second data set is built for this study
and consists of users who have active accounts both on Twitter and Flickr. Personality traits
for this group are estimated by analyzing their online text. Image interactions of these users on
both platforms are used in cross-modal and cross-platform analysis. Figs 1 and 2 show the pro-
cess of our cross-modal and cross-platform analysis. We also compare different features in pre-
dicting personality traits and perform experiments to uncover if cross-modal and cross-
platform data fusion can improve the predictive accuracy of personality. In the rest of this sec-
tion, we describe in more detail the data sets used in the analysis, the methods for obtaining
the features used in our results and the methodology for predicting personality traits.

Data
PsychoFlickr data set This data set contains a set of self-assessed and perceived personalities
for 300 random pro users from Flickr [47]. Pro users of Flickr are reportedly more likely to be

PLOS ONE | https://doi.org/10.1371/journal.pone.0198660 July 11, 2018 4 / 19

 Predicting user personality on Twitter and Flickr from posted, liked and profile images

Fig 2. Overview of cross-platform analysis.

https://doi.org/10.1371/journal.pone.0198660.g002

PLOS ONE | https://doi.org/10.1371/journal.pone.0198660 July 11, 2018 5 / 19

 Predicting user personality on Twitter and Flickr from posted, liked and profile images

Table 2. Descriptive statistics of the PsychoFlickr data set.

Flickr
Modality Total #images Average #images per user Median #images per user
Posts 72,997 247 170
Likes 60,001 203 200
Profile Images 295 1 1
https://doi.org/10.1371/journal.pone.0198660.t002

passionate about photography [47]. In this paper, we use the Flickr API (https://www.flickr.
com/services/api/) to extract profile images of those users and up to 300 of their posted and
liked images. We collect in total 295 profile images, 72,997 posted and 60,001 liked images in
this data set. Since we are interested in examining users’ personality, and not its perception, we
use the self-assessed personality traits in this study. Table 2 shows the descriptive statistics and
Fig 3 shows the distribution of different personality traits in this data set. We use this data set
for the cross-modal analysis—to compare user personality prediction across different modali-
ties (profile images, likes and posts).
Cross-linked Flickr-Twitter data set This data set contains a set of 334 users having both
Flickr and Twitter accounts and their estimated personality traits. All data was collected
according to Twitter and Flikr’s terms of service and privacy conditions. We do not have per-
sonality computed via surveys for this data set, as these are very costly and time-consuming to
administer. Hence, following previous work on personality analysis from profile images [33],
we use an automatic text-regression method to assign each user scores for the Big Five person-
ality traits [48]. The model was trained on a sample of over 70,000 users, using tokens and top-
ics extracted from status updates as features, achieving a validation predictive performance of r
*.35 on average for all five traits [48], which is considered a high correlation in psychology,
especially when measuring internal states [49]. For each user, we downloaded up to 3200 of
their most recent tweets using the Twitter API (https://dev.twitter.com/rest/public) to help
with predicting personality traits. Table 3 shows the descriptive statistics of this data set. We
use this data set for the cross-platform analysis.

Fig 3. Distribution of different personality traits at the two data sets. (a) Psycho-Flickr and (b) Cross-Linked Flickr and Twitter.
https://doi.org/10.1371/journal.pone.0198660.g003

PLOS ONE | https://doi.org/10.1371/journal.pone.0198660 July 11, 2018 6 / 19

 Predicting user personality on Twitter and Flickr from posted, liked and profile images

Table 3. Descriptive statistics of the Cross-Linked Flickr and Twitter data set.
Flickr
Modality Total # images Average # images per user Median #images per user
Posts 60,381 175 56
Likes 28,658 83 45
Profile Images 344 1 1
Twitter
Modality Total # images Average # images per user Median #images per user
Posts 73,576 213 199
Likes 29,030 84 82
Profile Images 344 1 1
https://doi.org/10.1371/journal.pone.0198660.t003

In building this data set, we selected the users who reported their Flickr profile in their
Twitter description. Further, we use the Flickr API to extract profile, and up to 300 posted and
liked images for each user, similar to the previous data set [47]. We collect a total of 334 profile
images, 60,381 posted and 28,657 liked images on Flickr.
For the same set of users, we collected image data using the Twitter API—a total of 334 pro-
file images, 73,576 posted and 61,197 liked images on Twitter. In order to obtain results com-
parable with the ones obtained on the Flickr data, we sub-sampled 29,030 liked images for the
analysis. Fig 3 shows the distribution of different personality traits.

Feature extraction
In order to study and interpret different modalities, we use two categories of features: colors
and content. The former contains basic color information and the latter contains information
extracted from the content of the image. For profile images, we use the features extracted from
the profile image of the user and for liked and posted images we perform a mean feature pool-
ing of all liked and posted images each across all images per user. The features used in this
study are summarized in Table 4 and are described below.

Table 4. Description of features used in this work.

Feature Type Dimension Feature Name Detailed Description
Color 1 Grayscale (binary) if an image is grayscale or not. If the image is grayscale, then the rest of the features are not
computed
10 HSV statistics Average and standard Deviation of hsv space, number of distinct hues, natural log of h_count
12 Hue statistics 12 hue histogram (normalized, all 12 values sum up to 1)
1 Pleasure (p) Pleasure = 0.69Brightness+0.22Saturation
1 Arousal (a) Arousal = 0.31Brightness+0.60Saturation
1 Dominance (d) Dominance = 0.76Brightness+0.32Saturation
6 6 Hue histogram yellow, green, cyan, blue, magenta, red
Content 1365 CNN object and scene VGG_Net prediction on 1000 objects and 365 scene categories
features probabilities
4096 CNN generic features 4096 dim penultimate layer features of VGG_Net
1299 Imagga tags list of Imagga tags for a set of images
https://doi.org/10.1371/journal.pone.0198660.t004

PLOS ONE | https://doi.org/10.1371/journal.pone.0198660 July 11, 2018 7 / 19

 Predicting user personality on Twitter and Flickr from posted, liked and profile images

Color Features Images are first converted to HSV space (Hue–Saturation–Value) as this
provides a more intuitive representation of colors for users [50]. A pixel in the HSV space is
characterized by three numbers: (1) Hue—the color type ranging between 0 and 360 degrees
e.g., 0 is red, 60 is yellow, is green; (2) Saturation—the intensity of the color ranging from 0 to
1 e.g., 0 means no color; (3) Value—the brightness of the color ranging from 0 to 1 e.g., 0 rep-
resents black. Using the HSV representation, we first divide images into grayscale and color
images. For color images, we calculate HSV statistics including mean and standard deviation
of hue, saturation and value. We extract brightness and saturation as the mean of saturation
and values respectively. An experimental study of colors established a linear relationship
between saturation and brightness and the dimensional model of affect containing three fac-
tors: Pleasure, Arousal and Dominance [51, 52]. We also extract the hue histogram count for
yellow, green, cyan, blue, magenta, red, the 12 color hue histogram, number of distinct hues
(h_count) and its natural log (log_h_count). Out of the 32 dimensional vector we extract,
some of these features have been applied in [47] to analyze personality of people who liked
images on the PsychoFlickr data set (Color).
Content Features To represent image content, we used features from convolutional neural
network trained on the Places data set [53] and tags derived using the convolutional neural
network based Imagga automatic image tagging system.
Convolutional networks (ConvNets or CNNs) have recently enjoyed a great success in
large-scale image recognition. A deep convolutional neural network architecture with 16-19
hidden layers named VGGnet is proposed in [53]. This classifier achieved the best results in
the ImageNet Large Scale Visual Recognition Challenge 2014 in the object classification and
localization challenge. We apply the proposed VGGnet model on our images. For a given
image xi, the last fully connected layer of the VGGnet—called the penultimate layer—produces
4096-dimensional activations, which are the high-level features used to represent image xi
(CNN_Gen). In addition, we apply the model trained on 1000 object ImageNet tagset [54] and
365 standard scene categories [55] and use the prediction probabilities across all scene and
object categories as features (CNN_Obj).
Images can have very diverse content beyond the ImageNet categories which have a limited
taxonomy relative to the content of social media images (e.g. not including faces or common
objects). We thus use the Imagga tagging API (http://docs.Imagga.com/#auto-tagging) as our
content analysis engine, which was successfully used by past research [56]. We labeled all
images with the Imagga Tagging API and generated for each image a bag-of-tags out of the
top-10 predicted tags, based on the developers’ recommendations. We removed all tags that
occurred less than 200 times in our largest data set, leaving us with 1,299 distinct tags.

Analysis
In our experiments, we first provide an analysis that shows how accurate each different set of
features is at the task of personality prediction. Then, we investigate which modality of interac-
tion—profile, posted or liked images—is most predictive of users’ personality. Finally, we
investigate which platform—Twitter or Flickr—is more predictive of users’ personality, and
how cross-platform fusion impacts prediction performance.
In all our experiments we use linear regression with Elastic Net regularization [57] as our
prediction algorithm. We tried L1 and L2 regularizers and Elastic Net regularization per-
formed better as they combine both L1 and L2 norms. Results are reported on 10-fold cross
validation measured by using Pearson correlation over the 10-folds. The same patterns hold
when evaluating the results with Root Mean Squared Error for regression and we omit them
for brevity. In all sections, feature combination is performed by training a linear ensemble

PLOS ONE | https://doi.org/10.1371/journal.pone.0198660 July 11, 2018 8 / 19

 Predicting user personality on Twitter and Flickr from posted, liked and profile images

over the individual prediction scores of each feature set. To test the significance of the models,
the F-statistic (ANOVA) and the p-values are reported. All the experiments are done repeat-
edly with randomized dataset splits for 100 times and standard deviation for all of the results
were found to be less than 0.001.

Results
In this section we answer the research questions raised in the Introduction.

Feature analysis
We compare the performance in predicting personality traits of the following sets of features,
across each modality: colors, CNN_Gen features (from the penultimate layer of VGGnet pro-
posed in [53]) and object/scene probabilities (from the final layer of the same network trained
on [55]) and Imagga tags. In addition, we build a model that uses a combination of the fea-
tures. Results are shown in Fig 4. The results show that:
(1) For profile images, CNN_Obj outperform other features for openness, conscientious-
ness, extraversion and neuroticism while CNN penultimate-layer features have the best perfor-
mance only for agreeableness. In this modality, color features and Imagga tags have similar
patterns. Their predictive performance is better for openness than conscientiousness and
agreeableness. We observe that combining features leads generally to better results.
(2) For posted images, Imagga tags and CNN penultimate-layer features generally achieve
the best predictive performance when compared to color and CCN categories. Imagga tags
and CNN penultimate-layer features have nearly the same predictive performance for consci-
entiousness, extraversion and agreeableness and CNN penultimate-layer features slightly out-
perform others for openness and neuroticism. The overall better accuracy of CNN
penultimate-layer features and Imagga tags is explainable by the fact that posted images con-
tain a very diverse array of objects and subjects—as opposed to profile pictures—which are
best captured by general image content features. CNN_Obj are not as good predictors proba-
bly due to the lack of diversity of the ImageNet categories, which do not include usual objects
and subjects encountered in social media images.

Fig 4. Prediction performance of models trained on different features: color, CNN generic features (CNN_Gen), CNN object and scene categories (CNN_Obj)
and Imagga tags; extracted from (a) profile images, (b) posted images and (c) liked images measured in Pearson correlation on the PsychoFlickr data set. All
Features denotes the performance of a model trained as linear ensemble of models trained on individual features. Significance of models is tested based on F-statistics
(ANOVA); +: p < 0.05,  : p < 0.01,  : p < 0.001.
https://doi.org/10.1371/journal.pone.0198660.g004

PLOS ONE | https://doi.org/10.1371/journal.pone.0198660 July 11, 2018 9 / 19

 Predicting user personality on Twitter and Flickr from posted, liked and profile images

(3) For liked images, Imagga and CNN penultimate-layer features achieve, similarly to
posted images, the best predictive results on all personality traits. CNN penultimate-layer fea-
tures outperform others in extraversion, neuroticism and openness, with the two achieving
similar performance on the other three traits. Again, in this modality, color and CNN_Obj fea-
tures follow similar patterns. Combining features does not add significantly to predictions on
each trait, which demonstrates that all feature types capture similar information.

Cross-modal analysis
In this experiment, we investigate the accuracy of models trained on different modalities at
predicting personality traits. Models are trained using all features extracted from images.
Results are presented in Fig 5 (showing a summarized view of results from Fig 4).
Profile pictures have the best performance in predicting conscientiousness and the lowest
in predicting agreeableness. Posted images have overall the best predictive performance, being
especially accurate at predicting agreeableness, openness to experience and neuroticism. This

Fig 5. Prediction performance of models trained on features extracted from profile, posted and liked images based on Pearson correlation on the
PsychoFlickr data set. Significance of models is tested based on F-statistics (ANOVA); +: p < 0.05,  : p < 0.01,  : p < 0.001.
https://doi.org/10.1371/journal.pone.0198660.g005

PLOS ONE | https://doi.org/10.1371/journal.pone.0198660 July 11, 2018 10 / 19

 Predicting user personality on Twitter and Flickr from posted, liked and profile images

is not unexpected as posted images are more than a single profile image and represent the
most direct way in which a user expresses his personality. Liked images do not achieve signifi-
cantly better results than any of the other modalities on any personality trait, but are on par
with posted images for neuroticism. Overall, this shows that liked images are not the most
direct way of expressing personality, while profile images are surprisingly effective in personal-
ity prediction, taking into account that this only represents a single image.

Cross-platform analysis
We investigate the predictive performance of images from two platforms in predicting differ-
ent personality traits using the Cross-Linked data set where we have collected posted, liked
and profile images from the same set of users on two different social media platforms: Twitter
and Flickr.
Psycho-Flickr consists of a set of users who answered the BFI survey [26] and Crossed-
Linked Flickr and Twitter consists of a set of users with active accounts both on Flickr and
Twitter. We used text mining approaches to predict personality traits for this set of users. To
examine the robustness of text-predicted labels, we train models on Cross-Linked Flickr and
Twitter data and test them on survey labeled personality traits of Psycho-Flickr dataset. We
binarise the labels on both datasets using quartile split (as done by Segalin et al. [12]), divide
the Cross-Linked Flickr and Twitter dataset into two splits (70% train and 30% test) for the
analysis. Baseline accuracy here is 50%. We apply a combination of Random Forests and Sup-
port Vector Machine classifiers that have been used in computer vision and social science
problems [58, 59]. The result are shown in Figs 6 to 8, where a comparison between models
trained on text-predicted labels and tested on survey labels versus models trained and tested
on questionnaire (survey) labels is presented. Models trained on text-predicted labels perform
as well as models trained on survey labels, if not better in some cases, perhaps due to the large
scale sample size used to train the text based model [28]. Further studies need to study this
behavior using large-scale survey based samples.
Results using a combination of all feature types are shown in Fig 9 for each modality (i.e.
profile pictures, posted and liked images) and platform, which we describe in more detail in
the following paragraphs.
Comparing Independent Modalities: For profile images, the results are largely similar,
with Flickr clearly outperforming Twitter only for neuroticism. For posted images, the perfor-
mance is relatively similar for all traits except conscientiousness where Flickr data achieves

Fig 6. Profile images. Comparison of models trained on text-predicted labels (Crossed-Link Flickr and Twitter) and those trained on survey label data at predicting
survey labels (Psycho-Flickr dataset) using (a) color features (b) CNN generic Features (c) CNN Probabilities on ImageNet Scene and Object Categories.
https://doi.org/10.1371/journal.pone.0198660.g006

PLOS ONE | https://doi.org/10.1371/journal.pone.0198660 July 11, 2018 11 / 19

 Predicting user personality on Twitter and Flickr from posted, liked and profile images

Fig 7. Posted images. Comparison of models trained on text-predicted labels (Crossed-Link Flickr and Twitter) and those trained on survey label data at predicting
survey labels (Psycho-Flickr dataset) using (a) color features (b) CNN generic Features (c) CNN Probabilities on ImageNet Scene and Object Categories.
https://doi.org/10.1371/journal.pone.0198660.g007

Fig 8. Liked images. Comparison of models trained on text-predicted labels (Crossed-Link Flickr and Twitter) and those trained on survey label data at predicting
survey labels (Psycho-Flickr dataset) using (a) color features (b) CNN generic Features (c) CNN Probabilities on ImageNet Scene and Object Categories.
https://doi.org/10.1371/journal.pone.0198660.g008

better performance. For liked images, Twitter data is overall most predictive than Flickr, with
the exception of openness to experience.
Comparing Combined Modalities: Next, we combine the three modalities (i.e. profile pic-
tures, posted and liked images) to see if we can improve predictive results, thus exploring if
these capture complimentary information and its extent. The results are shown in Fig 10. As
seen in this figure, combining modalities always results in better predictive performance and
in some cases, the improvement obtained is relatively large, for example in the case of open-
ness to experience and conscientiousness and Flickr. For Twitter, the improvements are rela-
tively smaller. This shows that in Flickr posting and liking images are more disparate actions,
while on Twitter their content is more similar.
Comparing Combined Platforms: Finally, we explore if combining information from both
platforms can result in an additional boost in prediction performance. We achieve this by
building a linear ensemble on top of the feature- and modality- ensembles build in the previ-
ous step. The results are shown in Fig 11. We can see that combining information from differ-
ent platforms (Flickr and Twitter) can additionally slightly improve results, with the exception
of extraversion. Overall, Flickr is more predictive of openness and conscientiousness, Twitter
is more predictive in case of extraversion, and for agreeableness and neuroticism, the

PLOS ONE | https://doi.org/10.1371/journal.pone.0198660 July 11, 2018 12 / 19

 Predicting user personality on Twitter and Flickr from posted, liked and profile images

Fig 9. Prediction performance of different platforms for (a) profile, (b) posted, (c) liked images based on Pearson correlation on Cross-Linked Twitter and Flickr
data set. Significance of models is tested based on F-statistics (ANOVA); +: p < 0.05,  : p < 0.01,  : p < 0.001.
https://doi.org/10.1371/journal.pone.0198660.g009

Fig 10. Prediction performance of combining different modalities (i.e. profile pictures, posted and liked images) versus using each modality separately on (a)
Flickr and (b) Twitter based on Pearson correlation on Cross-Linked Twitter and Flickr data set. Combined Modality denotes the performance of a model trained
as linear ensemble of models trained on individual modality. Significance of models is tested based on F-statistics (ANOVA); +: p < 0.05,  : p < 0.01,  : p < 0.001.
https://doi.org/10.1371/journal.pone.0198660.g010

performance is similar. In absolute terms, conscientiousness is most predictive overall, fol-

lowed by openness to experience and agreeableness. Extraversion is the least predictable per-
sonality trait.

Discussion
The results of this work confirm the hypothesis that multiple interactions that users have with
social media platforms such as choosing profile pictures, posting and liking images have pre-
dictive utility for automatic personality assessment of users, albeit with varying levels of perfor-
mance; and combining different interaction types and platforms, although it involves more
computation, can boost the prediction results. While posted images topped the performance

PLOS ONE | https://doi.org/10.1371/journal.pone.0198660 July 11, 2018 13 / 19

 Predicting user personality on Twitter and Flickr from posted, liked and profile images

Fig 11. Prediction performance for combining different modalities and different platforms based on Pearson correlation on Cross-Linked Twitter
and Flickr data set. Significance of models is tested based on F-statistics (ANOVA); +: p < 0.05,  : p < 0.01,  : p < 0.001.
https://doi.org/10.1371/journal.pone.0198660.g011

in predicting personality followed by liked images and then profile pictures, profile pictures
are a ubiquitous way for users to present themselves on social media, and they are usually con-
sidered public data which makes them easier to be accessed by automatic algorithms. Posted
and liked images, on the other hand, are relatively more diverse in their content and automatic
algorithms would need access to a larger set of such images across user’s posting timeline than
liked pictures to make accurate predictions.
Posted images specially perform well in predicting openness to experience which stands for
intellectual stimulation, willingness to explore new ideas, and similar traits. High prediction
performance using posted images can be associated to prior research [60] that has shown two
criteria for personality prediction to be successful—the environment in which users are must
allow them to express the trait (termed as ‘Relevance’) and, the trait must be perceptible to oth-
ers (termed as ‘Availability’), in this case automatic algorithms.

PLOS ONE | https://doi.org/10.1371/journal.pone.0198660 July 11, 2018 14 / 19

 Predicting user personality on Twitter and Flickr from posted, liked and profile images

Liking images can be a consequence of multiple motivations including social factors or

affective aspects such as reminding of positive events or ties with the people that have posted
them. ‘Likes’ are a way users publicly and voluntarily express appreciation for content online
[61]. As a result, users on several photo-sharing platforms create galleries of ‘favorite’ pictures
which provides computer vision and social science researchers a strong source of data for ana-
lyzing users personality.
Difference in the online social networking platforms also is an important consideration for
automatic methods to assess personality accurately. For example, Flickr is a social networking
site that is used by people who do photography more professionally [47]. Instead, Twitter is a
social media site on which users can share a diverse array of contents they are interested in
[62]. Comparing both platforms showed that Twitter performance is noticeably higher in pre-
dicting agreeableness from liked images. The fact that agreeable people tend to evaluate con-
tent favorably is represented in twitter more than Flickr. We also find that Flickr has a higher
performance at predicting conscientiousness from posted images, which corroborates the
hypothesis that Flickr is used by people who do photography more professionally.
A lot of systems can benefit from personality detection. For example, dating websites can try-
ing to match personalities of individuals before they meet each other [18]. Human Resources
department could predict job satisfaction before hiring a potential employee. Recommender
systems and commercial companies can improve their accuracy by recommending photos,
movies or music, that have higher chance to make positive impressions on their users. Knowl-
edge of a user’s personality also enables software developers to customize user interfaces [63].
This work provides multiple directions for future works. Psychological studies [64] show
that biological and socio-demographic factors are parameters in shaping an individual’s per-
sonality; thus, adding socio-demographic features such as ethnicity, language, cultural and
financial background, family size can potentially provide more insight. Further, it would be
interesting to study the information contained in social media usage which goes above and
beyond socio-demographics. While in this work we could not delve into providing insight due
to the restriction we had with the data size, future work on larger samples and developing
more interpretable visual features can serve to both boost performance [38] and to provide
more insight about the manifestation of personality online. Recently, methods based on Gauss-
ian Processes have been recently used to improve personality [65] and demographic [66] pre-
diction. Though they improve the state-of-the-art in user-trait modeling, they are unlikely to
significantly impact the answers to the research questions stated in our paper. We will leave
improving the prediction accuracy for future work.
The feasibility of social-media-based assessment of personality traits also raises ethical ques-
tions. Organizations with vested interests could exploit this information, for example, to
potentially influence people towards their agenda using social media. Data protection and
ownership frameworks are needed to make sure the data is not used against the users’ interest.
Few users realize the amount of psycho demographic information that can be gleaned from
their digital traces, so transparency about the derived indicators should be part of ethical and
policy discourse [67].

Conclusion
We carried out a cross-modal and cross-platform study using images posted on social media.
We used a wide range of color and semantic features extracted from images to analyze how dif-
ferent features can be applied to predict Big Five personality traits.
Posted images are generally more predictive than liked images and profile images, albeit
profile images obtained good results given that this only represents a single image. Overall,

PLOS ONE | https://doi.org/10.1371/journal.pone.0198660 July 11, 2018 15 / 19

 Predicting user personality on Twitter and Flickr from posted, liked and profile images

semantic features from CNN and Imagga are the best feature types for modelling the content
of posted and liked images.
Results on our novel cross-linked data set showed that Flickr provides overall better signal
than Twitter for predicting personality traits. Combining modalities is shown to generally
improve predictive performance especially in the case of Flickr, showing that the multiple
modalities encode more complimentary information than they do on Twitter. Finally, combin-
ing Flickr and Twitter information largely improves results, albeit not with wide margins.
Overall, our analysis shows that conscientiousness and openness to experience are the most
predictable personality traits from images posted online.

Author Contributions
Conceptualization: Zahra Riahi Samani.
Data curation: Zahra Riahi Samani, Sharath Chandra Guntuku, Daniel Preoţiuc-Pietro.
Methodology: Zahra Riahi Samani, Sharath Chandra Guntuku, Daniel Preoţiuc-Pietro.
Supervision: Sharath Chandra Guntuku, Lyle H. Ungar.
Visualization: Sharath Chandra Guntuku.
Writing – original draft: Zahra Riahi Samani, Sharath Chandra Guntuku.
Writing – review & editing: Sharath Chandra Guntuku, Mohsen Ebrahimi Moghaddam,
Daniel Preoţiuc-Pietro, Lyle H. Ungar.

References
1. Pew Research Center. Social Media Update 2016; 2016. http://www.pewinternet.org/2016/11/11/
social-media-update-2016/.
2. Damnianović V, Matović V, Kostić SC, Okanović M. The Role of the LinkedIn Social Media in Building
the Personal Image. Management (1820-0222). 2012;(65).
3. Honeycutt C, Herring SC. Beyond Microblogging: Conversation and Collaboration via Twitter. In: Pro-
ceedings of the 47th Hawaii International Conference on System Sciences. HICSS; 2009. p. 1–10.
4. Burdick A, Drucker J, Lunenfeld P, Presner T, Schnapp J. Digital Humanities; 2012.
5. Kosinski M, Stillwell D, Graepel T. Private Traits and Attributes are Predictable from Digital Records of
Human Behavior. Proceedings of the National Academy of Sciences. 2013; 110(15):5802–5805.
https://doi.org/10.1073/pnas.1218772110
6. Skowron M, Tkalčič M, Ferwerda B, Schedl M. Fusing Social Media Cues: Personality Prediction from
Twitter and Instagram. WWW Companion; 2016.
7. Manikonda L, Meduri VV, Kambhampati S. Tweeting the Mind and Instagramming the Heart: Exploring
Differentiated Content Sharing on Social Media. In: Proceedings of the Tenth International AAAI Confer-
ence on Weblogs and Social Media. ICWSM; 2016. p. 639–642.
8. Gosling SD, Ko SJ, Mannarelli T, Morris ME. A Room with a Cue: Personality Judgments based on
Offices and Bedrooms. Journal of Personality and Social Psychology. 2002; 82(3):379. https://doi.org/
10.1037/0022-3514.82.3.379 PMID: 11902623
9. Duggan M, Rainie L. Cell phone activities 2013. Pew Research Center’s Internet & American Life Proj-
ect Washington, DC; 2013.
10. Rainie L, Brenner J, Purcell K. Photos and videos as social currency online. Pew Internet & American
Life Project. 2012;.
11. Samani ZR, Moghaddam ME. A knowledge-based semantic approach for image collection summariza-
tion. Multimedia Tools and Applications. 2017; 76(9):11917–11939. https://doi.org/10.1007/s11042-
016-3840-1
12. Segalin C, Perina A, Cristani M, Vinciarelli A. The pictures we like are our image: continuous mapping of
favorite pictures into self-assessed and attributed personality traits. 2016;.
13. Furnham A, Jackson CJ, Miller T. Personality, learning style and work performance. Personality and
individual differences. 1999; 27(6):1113–1122. https://doi.org/10.1016/S0191-8869(99)00053-7

PLOS ONE | https://doi.org/10.1371/journal.pone.0198660 July 11, 2018 16 / 19

 Predicting user personality on Twitter and Flickr from posted, liked and profile images

14. Rentfrow PJ, Gosling SD. The do re mi’s of everyday life: The structure and personality correlates of
music preferences. Journal of personality and social psychology. 2003; 84(6):1236. https://doi.org/10.
1037/0022-3514.84.6.1236 PMID: 12793587
15. Guntuku SC, Ramsay JR, Merchant RM, Ungar LH. Language of ADHD in Adults on Social Media.
Journal of attention disorders. 2017; p. 1087054717738083. https://doi.org/10.1177/
1087054717738083 PMID: 29115168
16. Guntuku SC, Yaden DB, Kern ML, Ungar LH, Eichstaedt JC. Detecting depression and mental illness
on social media: an integrative review. Current Opinion in Behavioral Sciences. 2017; 18:43–49. https://
doi.org/10.1016/j.cobeha.2017.07.005
17. Preoţiuc-Pietro D, Eichstaedt J, Park G, Sap M, Smith L, Tobolsky V, et al. The role of personality, age,
and gender in tweeting about mental illness. In: Proceedings of the 2nd Workshop on Computational
Linguistics and Clinical Psychology: From Linguistic Signal to Clinical Reality; 2015. p. 21–30.
18. Kaushal V, Patwardhan M. Emerging Trends in Personality Identification Using Online Social Networks
—A Literature Survey. ACM Transactions on Knowledge Discovery from Data (TKDD). 2018; 12(2):15.
https://doi.org/10.1145/3070645
19. Komarraju M, Karau SJ. The relationship between the big five personality traits and academic motiva-
tion. Personality and individual differences. 2005; 39(3):557–567. https://doi.org/10.1016/j.paid.2005.
02.013
20. Guntuku SC, Scott MJ, Ghinea G, Lin W. Personality, Culture, and System Factors-Impact on Affective
Response to Multimedia. arXiv preprint arXiv:160606873. 2016;.
21. Guntuku SC, Lin W, Scott MJ, Ghinea G. Modelling the influence of personality and culture on affect
and enjoyment in multimedia. In: Affective Computing and Intelligent Interaction (ACII), 2015 Interna-
tional Conference on. IEEE; 2015. p. 236–242.
22. Scott MJ, Guntuku SC, Lin W, Ghinea G. Do personality and culture influence perceived video quality
and enjoyment? IEEE Transactions on Multimedia. 2016; 18(9):1796–1807. https://doi.org/10.1109/
TMM.2016.2574623
23. Guntuku SC, Scott MJ, Yang H, Ghinea G, Lin W. The CP-QAE-I: A video dataset for exploring the
effect of personality and culture on perceived quality and affect in multimedia. In: Quality of Multimedia
Experience (QoMEX), 2015 Seventh International Workshop on. IEEE; 2015. p. 1–7.
24. Zhu Y, Guntuku SC, Weisi L, Ghinea G, Redi JA. Measuring Individual Video QoE: A Survey, and Pro-
posal for Future Directions Using Social Media. ACM Transactions on Multimedia Computing, Commu-
nications and Applications. 2018;.
25. Costa PT, MacCrae RR. Revised NEO Personality Inventory (NEO PI-R) and NEO Five-factor Inven-
tory (NEO FFI): Professional Manual. Psychological Assessment Resources; 1992.
26. McCrae RR, John OP. An Introduction to the Five-factor Model and its Applications. Journal of Person-
ality. 1992; 60. https://doi.org/10.1111/j.1467-6494.1992.tb00970.x
27. Celli F, Pianesi F, Stillwell D, Kosinski M. Workshop on Computational Personality Recognition (Shared
Task). In: Proceedings of the Workshop on Computational Personality Recognition (WCPR). ICWSM;
2013.
28. Schwartz HA, Eichstaedt JC, Kern ML, Dziurzynski L, Ramones SM, Agrawal M, et al. Personality, Gen-
der, and Age in the Language of Social Media: The Open-vocabulary Approach. PloS ONE. 2013; 8(9).
https://doi.org/10.1371/journal.pone.0073791
29. Ferwerda B, Schedl M, Tkalčič M. Using Instagram Picture Features to Predict Users’ Personality.
MMM; 2016.
30. Celli F, Bruni E, Lepri B. Automatic Personality and Interaction Style Recognition from Facebook Profile
Pictures. In: Proceedings of the 22nd ACM international conference on Multimedia. MM; 2014. p. 1101–
1104.
31. Nie J, Cui P, Yan Y, Huang L, Li Z, Wei Z. How your Portrait Impresses People?: Inferring Personality
Impressions from Portrait Contents. In: Proceedings of the 22nd ACM International Conference on Mul-
timedia. MM; 2014. p. 905–908.
32. Al Moubayed N, Vazquez-Alvarez Y, McKay A, Vinciarelli A. Face-Based Automatic Personality Per-
ception. MM; 2014.
33. Liu L, Preotiuc-Pietro D, Samani ZR, Moghaddam ME, Ungar LH. Analyzing Personality through Social
Media Profile Picture Choice. In: Proceedings of the Tenth International AAAI Conference on Weblogs
and Social Media. ICWSM; 2016. p. 211–220.
34. Ferwerda B, Tkalcic M. You Are What You Post: What the Content of Instagram Pictures Tells About
Users’ Personality. In: The 23rd International on Intelligent User Interfaces; 2018.
35. Nie J, Wei Z, Li Z, Yan Y, Huang L. Understanding personality of portrait by social embedding visual fea-
tures. Multimedia Tools and Applications. 2018; p. 1–20.

PLOS ONE | https://doi.org/10.1371/journal.pone.0198660 July 11, 2018 17 / 19

 Predicting user personality on Twitter and Flickr from posted, liked and profile images

36. Guntuku SC, Lin W, Carpenter J, Ng WK, Ungar LH, Preoţiuc-Pietro D. Studying personality through
the content of posted and liked images on Twitter. In: Proceedings of the 2017 ACM on web science
conference. ACM; 2017. p. 223–227.
37. Guntuku SC, Qiu L, Roy S, Lin W, Jakhetiya V. Do Others Perceive You As You Want Them To?:
Modeling Personality based on Selfies. In: Proceedings of the 1st International Workshop on Affect &
Sentiment in Multimedia. ACM; 2015. p. 21–26.
38. Guntuku SC, Roy S, Weisi L. Personality modeling based image recommendation. In: International
Conference on Multimedia Modeling. Springer; 2015. p. 171–182.
39. Guntuku SC, Zhou JT, Roy S, Weisi L, Tsang IW. Who likes What, and Why? Insights into Personality
Modeling based on ImageLikes’. IEEE Transactions on Affective Computing. 2016;.
40. Wei H, Zhang F, Yuan NJ, Cao C, Fu H, Xie X, et al. Beyond the words: Predicting user personality from
heterogeneous information. In: Proceedings of the tenth ACM international conference on web search
and data mining. ACM; 2017. p. 305–314.
41. Nie J, Huang L, Cui P, Li Z, Yan Y, Wei Z, et al. Social Media Profiler: Inferring Your Social Media Per-
sonality from Visual Attributes in Portrait. In: Pacific Rim Conference on MM; 2016.
42. Sang J, Zhang H, Xu C. Visual BFI: An Exploratory Study for Image-Based Personality Test. PCM;
2016.
43. Segalin C, Cheng DS, Cristani M. Social Profiling through Image Understanding: Personality Inference
using Convolutional Neural Networks. Computer Vision and Image Understanding. 2017; 156:34–50.
https://doi.org/10.1016/j.cviu.2016.10.013
44. Segalin C, Perina A, Cristani M, Vinciarelli A. The pictures we like are our image: continuous mapping of
favorite pictures into self-assessed and attributed personality traits. IEEE Transactions on Affective
Computing. 2017; 8(2):268–285. https://doi.org/10.1109/TAFFC.2016.2516994
45. Segalin C, Celli F, Polonio L, Kosinski M, Stillwell D, Sebe N, et al. What your Facebook profile picture
reveals about your personality. In: Proceedings of the 2017 ACM on Multimedia Conference. ACM;
2017. p. 460–468.
46. Xiong X, Filippone M, Vinciarelli A. Looking Good With Flickr Faves: Gaussian Processes for Finding
Difference Makers in Personality Impressions. MM; 2016.
47. Cristani M, Vinciarelli A, Segalin C, Perina A. Unveiling the multimedia unconscious: Implicit cognitive
processes and multimedia content analysis. In: ACM MM; 2013.
48. Park G, Schwartz HA, Eichstaedt JC, Kern ML, Kosinski M, Stillwell DJ, et al. Automatic Personality
Assessment through Social Media Language. Journal of Personality and Social Psychology. 2014; 108
(6):934–952. https://doi.org/10.1037/pspp0000020 PMID: 25365036
49. Meyer GJ, Finn SE, et al. Psychological Testing and Psychological Assessment: A Review of Evidence
and Issues. American Psychologist. 2001;.
50. Bigun J. Vision with Direction: A Systematic Introduction to Image Processing and Vision; 2006.
51. Mehrabian A, Russell JA. An Approach to Environmental Psychology. MIT Press; 1974.
52. Russell JA, Mehrabian A. Evidence for a Three-Factor Theory of Emotions. Journal of Research in Per-
sonality. 1977; 11(3). https://doi.org/10.1016/0092-6566(77)90037-X
53. Simonyan K, Zisserman A. Very Deep Convolutional Networks for Large-Scale Image Recognition.
CoRR;.
54. Russakovsky O, Deng J, Su H, Krause J, Satheesh S, Ma S, et al. Imagenet large scale visual recogni-
tion challenge. International Journal of Computer Vision. 2015; 115(3):211–252. https://doi.org/10.
1007/s11263-015-0816-y
55. Zhou B, Khosla A, Lapedriza A, Torralba A, Oliva A. Places: An Image Database for Deep Scene
Understanding. arXiv preprint arXiv:161002055. 2016;.
56. Garimella VRK, Alfayad A, Weber I. Social Media Image Analysis for Public Health. In: Proceedings of
the 2016 CHI Conference on Human Factors in Computing Systems. CHI; 2016. p. 5543–5547.
57. Zou H, Hastie T. Regularization and variable selection via the elastic net. Journal of the Royal Statistical
Society: Series B (Statistical Methodology). 2005; 67(2):301–320. https://doi.org/10.1111/j.1467-9868.
2005.00503.x
58. Abdollahpour Z, Samani ZR, Moghaddam ME. Image classification using ontology based improved
visual words. In: Electrical Engineering (ICEE), 2015 23rd Iranian Conference on. IEEE; 2015. p. 694–
698.
59. Palomino-Garibay A, Camacho-Gonzalez AT, Fierro-Villaneda RA, Hernandez-Farias I, Buscaldi D,
Meza-Ruiz IV, et al. A random forest approach for authorship profiling. In: Proceedings of CLEF; 2015.

PLOS ONE | https://doi.org/10.1371/journal.pone.0198660 July 11, 2018 18 / 19

 Predicting user personality on Twitter and Flickr from posted, liked and profile images

60. Wright AG. Current directions in personality science and the potential for advances through computing.
IEEE Transactions on Affective Computing. 2014; 5(3):292–296. https://doi.org/10.1109/TAFFC.2014.
2332331
61. Ling K, Beenen G, Ludford P, Wang X, Chang K, Li X, et al. Using social psychology to motivate contri-
butions to online communities. Journal of Computer-Mediated Communication. 2005; 10(4):00–00.
https://doi.org/10.1111/j.1083-6101.2005.tb00273.x
62. Jaika K, Guntuku SC, Ungar LH. Facebook vs. Twitter: Cross-platform Differences in Self-disclosure
and Trait Prediction. In: ICWSM; 2018.
63. Guntuku SC, Roy S, Lin W, Ng K, Keong NW, Jakhetiya V. Personalizing User Interfaces for improving
quality of experience in VoD recommender systems. In: Quality of Multimedia Experience (QoMEX),
2016 Eighth International Conference on. IEEE; 2016. p. 1–6.
64. Wei X, Stillwell D. How Smart Does Your Profile Image Look?: Estimating Intelligence from Social Net-
work Profile Images. In: Proceedings of the Tenth ACM International Conference on Web Search and
Data Mining. ACM; 2017. p. 33–40.
65. Arnoux PH, Xu A, Boyette N, Mahmud J, Akkiraju R, Sinha V. 25 Tweets to Know You: A New Model to
Predict Personality with Social Media. arXiv preprint arXiv:170405513. 2017;.
66. Preoţiuc-Pietro D, Lampos V, Aletras N. An analysis of the user occupational class through Twitter con-
tent. In: Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and
the 7th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). vol.
1; 2015. p. 1754–1764.
67. McKee R. Ethical issues in using social media for health and health care research. Health Policy. 2013;
110(2):298–301. https://doi.org/10.1016/j.healthpol.2013.02.006 PMID: 23477806

PLOS ONE | https://doi.org/10.1371/journal.pone.0198660 July 11, 2018 19 / 19