Profilio: Psychometric Profiling to Boost Social Media Advertising

Abstract

Profilio is a profiling solution for the field of paid social media advertising. In particular the solution is designed for predictive data enrichment of Customer Relationship Management databases and for the segmentation of customer audiences. Three different Proof of Concepts with different clients have showed that the solution reduces the costs of paid social media advertising in different settings and with different advertising targets, especially starting from large audiences. In this paper we report the details about Profilio's technologies and the results of the Proof of Concepts.

Full text

Profilio: Psychometric Profiling to Boost Social Media
Advertising
Fabio Celli
Prolio Company
Via Sommarive 18
Trento, Italy 38123
fabio.celli@prolio.co
Pietro Zani Massani
Prolio Company
Via Sommarive 18
Trento, Italy 38123
pietro.zani@prolio.co
Bruno Lepri
FBK
Via Sommarive 18
Trento, Italy 38123
lepri@k.eu
ABSTRACT
Prolio is a proling solution for the eld of paid social media ad-
vertising. In particular the solution is designed for predictive data
enrichment of Customer Relationship Management databases and
for the segmentation of customer audiences. ree dierent Proof
of Concepts with dierent clients have showed that the solution
reduces the costs of paid social media advertising in dierent set-
tings and with dierent advertising targets, especially starting from
large audiences. In this paper we report the details about Prolio’s
technologies and the results of the Proof of Concepts.
CCS CONCEPTS
•Human-centered computing →User models;
Collaborative
and social computing;
KEYWORDS
User Proling, Personality Computing, Social Media, Advertising,
CRMs
ACM Reference format:
Fabio Celli, Pietro Zani Massani, and Bruno Lepri. 2017. Prolio: Psychome-
tric Proling to Boost Social Media Advertising. In Proceedings of MM’17,
October 23–27, 2017, Mountain View, CA, USA., , 5 pages.
DOI: hps://doi.org/10.1145/3123266.3129311
1 MOTIVATION AND BACKGROUND
Popular social media platforms such as Facebook, Twier, LinkedIn
and YouTube are more and more oering novel ways to advertise
brands. For example, Facebook provides to the advertisers options
such as promoted posts, sponsored stories, page post ads, etc. More-
over, Facebook, Twier and LinkedIn have developed a targeting
technology which allows advertisers to reach a specic audience
(e.g. males vs. females, people with specic interests or living in
specic places, dierent age groups, etc.). ese advanced targeting
options provide a level of personalization not achievable on other
advertising channels (e.g. TV, newspapers, websites, etc.). For ex-
ample, advertisers may reach specic audiences by looking at their
self-reported interests, skills, specic pages they are engaged with,
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
for prot or commercial advantage and that copies bear this notice and the full citation
on the rst page. Copyrights for components of this work owned by others than ACM
must be honored. Abstracting with credit is permied. To copy otherwise, or republish,
to post on servers or to redistribute to lists, requires prior specic permission and/or a
fee. Request permissions from permissions@acm.org.
MM’17, October 23–27, 2017, Mountain View, CA, USA.
©2017 ACM. ISBN 978-1-4503-4906-2/17/10. . . $15.00
DOI: hps://doi.org/10.1145/3123266.3129311
and so on. e interest targeted by this strategy, called interest
targeting, can be as general as an industry (e.g. fashion industry)
or as specic as a product (e.g. sunglasses). Currently, Facebook,
Twier, and LinkedIn are oering this targeting strategy.
Other examples of targeting strategies are behavioral targeting,
where an advertiser can reach specic customers based on their
purchase behaviors [
22
,
31
] and connection targeting, where an ad-
vertiser can reach people who have a specic kind of connection to
a given page, app, event or group. Both types of targeting, currently
oered by Facebook, LinkedIn and Twier, take customers’ past
behavior into account.
Finally, Facebook and Twier oer custom targeting, that per-
mits advertisers to reach specic audiences by uploading a list of
email addresses, phone numbers, usernames or users IDs
1
; while
Facebook and LinkedIn oer lookalike targeting, a strategy that
permits to reach new people who are similar to an audience of in-
terest. More specically, custom targeting permits to an advertiser
to upload and target directly an already known group of people.
Instead, lookalike targeting helps companies extend their custom
audiences to reach new, similar users. us, for those businesses
looking to acquire new customers through social media advertising,
lookalike targeting may work as a very eective acquisition tool.
Given all these targeting strategies, companies have doubled
social media advertising budgets worldwide over the past 2 years,
going from
$
16 billions in 2014 to
$
31 billions in 2016 [
3
]. Specif-
ically, paid social media advertising is primarily used to support
branding-related eorts, such as the consumer’s ability to recognize
or recall a brand, brand awareness, that is central to purchasing de-
cision making [
14
,
17
]. Let us take as example the fashion industry.
According to the McKinsey Global Fashion Index, this industry is
now worth about
$
2.4 trillion; however 2016 was one of the fashion
industry’s toughest years [
4
]. is is due to many factors such
as the competition of the emerging markets like China and India,
the stagnating economies in the Western countries, the increasing
volatility and speed of the market, and the growth of athletic wear.
At the same time, consumers have become more demanding and
less predictable in their purchasing behavior, and this translates
into higher advertising costs for brand awareness. us, this situa-
tion requires more powerful technological solutions for proling
and targeting consumers.
Crucially, consumers are nowadays leaving a lot of digital foot-
prints (e.g. Facebook prole pictures, tweets, Facebook likes and
statuses, pictures on Instagram, videos on YouTube, networks of ac-
quaintances on LinkedIn and of friends on Facebook, etc.) that can
1
Facebook calls its own audiences custom audiences, while Twier calls its own ones
tailored audiences

MM’17, , October 23–27, 2017, Mountain View, CA, USA. Fabio Celli, Pietro Zani Massani, and Bruno Lepri
be exploited to analyze and predict their behaviors, interests, politi-
cal and sexual orientations, and psychometric characteristics (e.g.
personality traits, emotional dispositions, etc.). Prolio Company,
launched at beginning of March 2017, has therefore developed a
proling technology, based on these dierent (multimodal) sources
of digital footprints (e.g. images, textual contents, demographics,
social media activities, etc.), that is able to predict a set of behavioral
variables, including purchase motivations, job performance and
subjective well-being. ere are many potential applications of this
technology, ranging from credit scoring to human resources, but
we aim to apply it to the market of paid social media advertising,
starting vertically from the eld of fashion and then expanding
to other elds. In Section 2 we describe our business idea, our
technology and the science behind, in Section 3 we report some
results obtained with our current customers and in Section 4 we
draw our conclusions and our perspective for the future.
2 BUSINESS IDEA
Prolio Company is the result of years of research on psychome-
tric proling and automatic behavior understanding from multi-
modal sources of data (e.g. text, images, social media activities,
ego-network characteristics, etc.). Specically, Prolio Company
oers a technological solution that responds to the growing demand
of understanding and predicting consumers’ behaviors. Starting
from public digital footprints, like social media prole pictures and
text, and a growing number of other data sources including demo-
graphics, Prolio Company’s psychometric engine predicts a set of
behavioral dimensions such as:
•
purchase behavior (aention to advertising, impulsive or
compulsive buyers, high spending customers),
•
purchase motivations (probability to seek for a sense of
belonging to a brand or to display a status),
•
cultural aitude (aitude towards innovation or conserva-
tion),
•
job performance (ability to manage stress, individual or
group task prociency, leading ability),
•
subjective and physical well-being (life satisfaction, health
probability),
•
relationships (anxiety, avoidance and relationship quality),
•
news sharing (probability to share news with negative
mood or to share fake news),
•
music preference (probability to like complex, dance, rebel
or conventional music),
•
tourism aitude (probability to seek adventure in travel
and to experience satisfaction from hospitality),
•
social aitude (probability to have pro-social or antisocial
behavior, aractiveness and social views).
We deliver this technological solution in a scalable way by means
of APIs and we provide a proling service to segment and create
custom audiences. e value of our technological solution lies in:
•
the enrichment of Customer Relationship Managements’
(CRM) customer data, that can be exploited for business
intelligence (i.e. customer segmentation),
•
the production of custom audiences that reduce the costs
of paid social media advertising.
e development of this solution and the business idea is based on
two pillars: a solid scientic work behind our technology and a
competitive advantage.
2.1 e Science Behind our Solution
Our technology is based on well-established theories in social psy-
chology and on a decade of research in the elds of personality
computing and human behavior understanding.
Research in social psychology has proposed and validated the-
ories to predict and explain individual behaviors and preferences
with psychological models, such as the Myers Briggs model [
19
].
For our solution we adopted the most widely accepted personality
model in the scientic community: the Five Factor Model, that de-
nes ve traits, namely Openness, Conscientiousness, Extroversion,
Agreeableness and Emotional Stability (oen conversely referred
to as Neuroticism) [10].
Research made great progresses in recent years by exploiting big
data of digital footprints from social media as an excellent ground
for personality computing and human behavior understanding. For
example, nowadays it is possible to train machine learning models
to predict personality types from Twier, Facebook or Linkedin,
using features such as number of followers, density of subject’s
network, number of hashtags, Facebook Likes and other language
independent features [9, 11, 12, 15, 16, 21, 23, 29].
An interesting nding is that the computational models based on
the subjects’ interests are signicantly more correlated with person-
ality self-ratings than average human judgments [
33
]. Sometimes
the availability of these features could be subject to limitations due
to privacy seings, but more recent work showed how it is possi-
ble to predict personality types from public prole pictures [
8
,
26
]
exploiting techniques such as Bag-of-Visual-Words, Convolutional
Neural Networks or other low level features [
7
,
27
,
32
], that are
able to capture information from technical variables of images such
as pixel paerns, color scales or variations in brightness. Prole
pictures (not necessarily faces) convey a lot of information about
a user and, according to literature in psychology, are directly con-
nected to their identity [
2
,
30
]. For example, results in this eld
revealed that extroverted and emotionally stable people tend to
have pictures where they are smiling and appear with other people;
introverts tend to appear alone and with less bright colors; neu-
rotics tend to have images with close-up faces or without humans
and uncooperative people tend to have pictures with few colors.
In addition to these ndings there is a rich scientic literature
about behavioral dimensions correlated to personality, such as so-
cioeconomic status, life satisfaction, job performance, relationship
quality and cognitive ability [
13
,
20
,
24
]. e accuracy of the system,
evaluated with training/test split and backtests for all the predicted
dimensions, is around 80%.
Exploiting data from Facebook and Twier collected since 2014
within a research project [
8
], we put together all these scientic
ndings in a customized system involving semi-supervised [
34
] and
Deep Learning techniques (e.g. Convolutional Neural Networks)
[
5
], beside other machine learning algorithms such as Support
Vector Machine regressors [
28
] and enhanced rule-based Ripper
algorithms [
25
]. A representation of the system is depicted in Figure
1.

Profilio: Psychometric Profiling to Boost Social Media Advertising MM’17, , October 23–27, 2017, Mountain View, CA, USA.
Figure 1: A schematic representation of Proling technol-
ogy
e system can take as input social media prole pictures (blurred
in Figure 1 for privacy reasons), text or demographic information.
A rst level of technical feature extraction feeds several dierent
layers of models that output the nal predictions and delivers a user
prole. For business reasons we can not disclose other information
about the system.
As our technological solution produces predictions of individ-
ual psychometric traits, we must be compliant on two main legal
aspects:
•
User consent on data treatment. For this reason, we use
exclusively public data as sources for our predictions (e.g.
prole pictures on Facebook, textual content on Twier,
etc.) or data owned/authorized by our clients (e.g. email
addresses and demographics), thus they bear the responsi-
bility for that data;
•
Discrimination. For this reason, our technological solution
avoids to produce discriminative predictions regarding po-
litical views, religious views, and sexual orientations.
2.2 Market, Competitors and Competitive
Advantage
Prolio’s reference market is divided into two main segments:
•
sales of data enrichment services for Customer Relation-
ship Managements (CRMs) of e-commerce portals (exam-
ples of target in the eld of fashion are Zalando and Yoox-
Net-A-Porter) and large consulting rms (e.g. Bain & Com-
pany, McKinsey, etc.)
•
sales of custom audiences proling for paid social media
advertising to large groups in the eld of fashion and lux-
ury (Kering holding who owns Gucci, Puma, and Yves
Saint Laurent among the other brands, Prada group, AEFFE
group, etc.)
e business model envisages two sources of income linked to each
other:
•
income from sales of proling services for data enrichment
and custom audiences’ generation,
•
income from research projects nalized to improve the
proling technologies.
e solutions available from the major global competitors are
general purpose products (with the exception of Cambridge Ana-
lytica that is focused on electoral processes) and mainly designed
for the US market. Apply Magic Sauce
2
is a company that provides
an API service predicting personality and other personal dimen-
sions from Likes and text in social media, sold with two types of
subscriptions: basic (
$
500 / month) and pro (
$
3000 / month). Cam-
bridge Analytica
3
is a company that combines data mining and
psychometric proling with strategic communication for electoral
processes. e company is heavily funded by the family of Robert
Mercer, an American hedge-fund billionaire [
1
]. In 2015 it became
known as the data analysis company working for Ted Cruz and
Donald Trump’s presidential campaigns. IBM Watson Personality
Insights
4
is a personality prediction service (metrics and summary
on personality and other business dimensions) delivered through
APIs. Prediction is made from various textual data (email, social
media, generic texts) in a limited number of languages at a price of
about
$
0.01 per call. Crystal
5
is a service that predicts 4 dimensions
of digital footprint personality with the DISC (Dominance, Induce-
ment, Submission, and Compliance) model [
18
] and specializes in
the sales sectors (predicts the most suitable loyalty to convince the
customer based on the personality prole) and hiring (predicts the
characteristics of candidates in the human resources sector) start-
ing from LinkedIn, email and calendar data. Pricing and service
delivery are customized.
In the reference market, knowledge acquisition from digital foot-
prints is hampered by a number of factors:
•
social media privacy seings, that prevent the collection
of some private kind of data, such as text and likes. e
solution oered by Prolio Company includes proling
from public sources like prole pictures.
•
limitations in the availability of data sources: many large
companies, even in the eld of luxury and fashion, have
2hp://applymagicsauce.com/
3hps://cambridgeanalytica.org
4hps://watson-pi-demo.mybluemix.net
5hps://www.crystalknows.com

MM’17, , October 23–27, 2017, Mountain View, CA, USA. Fabio Celli, Pietro Zani Massani, and Bruno Lepri
only demographic data of their customers in their CRMs,
and this prevents the application of many proling models;
Prolio Company oers a solution that is domain-adaptive
and is constantly concentrating the eorts towards enlarg-
ing the number of data sources that can be proled.
•
language limitations, that is especially important in a mar-
ket like EU. Prolio’s technology is language independent
and can be applied in dierent language areas.
ese factors, plus the fact that the application of these technolo-
gies in the eld of fashion is very recent, represent a competitive
advantage for Prolio Company.
3 CASE HISTORY AND RESULTS
Since March 2017, Prolio Company has ran three dierent Proof
of Concepts (PoC) with dierent clients, (i) a big fashion group,
(ii) a small startup promoting a discount app and (iii) a medium
e-commerce of fashion, design and food & wine. In each PoC
we tested the results of same advertising copyright with dierent
audiences in a span of time of one week.
Table 1: results of three dierent Proof of Concepts with
three dierent clients in a time span of one week.
PoC 1: big fashion group, target: video brand awareness
audience lookalike country CPV CTR Avg%ViewTime
 fans 3% UK 0.14 1% 16.6%
purchase 3% UK 0.04 24.4% 46.5%
Prolio 1 3% UK 0.04 24.6% 48.6%
PoC 2: small startup, target: app install
audience lookalike country CPA CPuC CPShare
 fans 3% IT 0.46 0.18 26.1
Prolio 2 3% IT 0.43 0.16 19.7
PoC 3: e-commerce, target: lead generation
audience lookalike country CPL CPuC CPShare
 fans 1% IT 1.12 0.18 115.8
subscribe 1% IT 1.02 0.26 262
purchase 1% IT 0.89 0.23 197.7
Prolio 3 1% I T 0.77 0.17 76.5
e results reported in Table 1 refer to the three clients in a time
span of one week. e target is dierent for each client: in PoC 1
(big fashion group) the target is brand awareness with a 15 seconds
viral videoclip. As evaluation metrics we used Cost Per 3 seconds
View (CPV), Click rough Rate (CTR) and average percentage of
videoclip view time (Avg%ViewTime). Starting from the fashion
group’s Facebook fans, we created a custom audience (Prolio
1) with users that our system predicted having high aention to
advertising. We compared this audience against two control groups:
(i) the Facebook fanbase of the fashion group and (ii) an audience
with users that purchased items of the fashion group. In order
to have audiences comparable also by potential reach, we created
Facebook lookalike of all the audiences in order to reach 1.4M
potential customers in UK with each audience. Our results show
that Prolio Company’s proling solution helps reducing CPV and
improving CTR and the average percentage of videoclip view time.
It is interesting to note that proling from a very large Facebook
fanbase obtains beer results than using an audience that purchased
brand’s items, at the same cost per view. In PoC 2 (small startup) the
target is an App install action. Again, we created a custom audience
(Prolio 2) starting from the startup’s Facebook fans, proling users
with high aention to advertising. We compared the results against
a control group with the startup Facebook fans, creating lookalike
audiences in order to have a comparable potential reach (150K
potential customers). Our results show that the proled audience
reduces the Cost per App Install (CPA), the Cost Per unique Clicks
(CPuC) and the Cost Per Shares. Finally, in PoC 3 (medium e-
commerce) the target is lead generation, dened as the registration
to the e-commerce website. Hence, we used as evaluation metrics
the Cost Per Lead (CPL), the Cost Per unique Clicks (CPuC) and
the Cost Per Shares. In this case, we have started creating our
proled audience from a mixture of the e-commerce Facebook fans,
Facebook users who reacted in the e-commerce Facebook page and
subscribed users. Again, we created lookalike audiences in order
to have a comparable potential reach (in this case 300K potential
customers) between experimental and control audiences. Results
show that our proling helps reducing CPL, CPuC and CPShare.
4 CONCLUSION AND FUTURE
Prolio Company is a startup in its early business development
stage. We have developed a Minimal Viable Product (MVP) with
a psychometric engine as its core. e solution, that delivers the
enrichment of CRMs’ data and the proling of custom audiences, re-
duces the costs of paid social media advertising in dierent seings
and with dierent advertising targets. e Proof Of Concepts we
ran with our rst customers showed that our proling solution is
eective, especially starting from large audiences. is is one of the
reasons why we target large fashion groups as our ideal clients. We
have several targets for the future of Prolio Company. First of all,
we are working for improving the performance of our psychometric
engine by enlarging the number of social media we use for training
our models, as scientic literature suggests that rich personas can
be extracted by looking at dierent social media as dierent points
of view on customers’ proles [
6
]. We are also working to improve
machine learning models by leveraging state-of-the-art approaches
in deep learning. Another goal is to automate also the generation
of advertising strategies customized for the specic psychometric
characteristics of our clients’ customers. More specically, we plan
to investigate the advertising eectiveness of generating targeted
multimodal (textual, visual and audio) messages. From a business
point of view, we are working to nd new clients, to explore new
markets (e.g. from Italy to Europe and then United States), and to
nd investors eager in taking the company to the next level. As
people are producing always larger samples of digital footprints
and privacy becomes more and more a serious issue, predictive pro-
ling will become more and more useful for business companies.
Prolio Company is working in this direction, with the mission to
automatically scouting the value of people behaviors, and turn it
into knowledge.
REFERENCES
[1]
2015. Cruz-Connected Data Miner Aims to Get Inside U.S. Vot-
ers’ Heads. hps://www.bloomberg.com/politics/features/2015-11- 12/
is-the- republican-party- s-killer-data- app-for- real-. (2015). Retrieved: February
2016.
[2] 2016. Analyzing Personality through Social Media Prole Picture Choice.

Profilio: Psychometric Profiling to Boost Social Media Advertising MM’17, , October 23–27, 2017, Mountain View, CA, USA.
[3]
2016. Social media advertising usage statistics. hps://blog.hootsuite.com/
social-media- advertising-stats. (2016). Retrieved: May 2017.
[4]
2017. e state of fashion. hp://www.mckinsey.com/industries/retail/
our-insights/the- state-of-fashion. (2017). Retrieved: May 2017.
[5]
Yoshua Bengio. 2009. Learning deep architectures for AI. Foundations and trends
®
in Machine Learning 2, 1 (2009), 1–127.
[6]
Kseniya Buraya, Aleksandr Farseev, Andrey Filchenkov, and Tat-Seng Chua. 2017.
Towards user personality proling from multiple social networks. In Proceedings
of the irty-First AAAI Conference on Articial Intelligence.
[7]
Fabio Celli. 2015. IATO: Feature Extraction for image analysis. Technical Report.
University of Trento.
[8]
Fabio Celli, Elia Bruni, and Bruno Lepri. 2014. Automatic Personality and Inter-
action Style Recognition from Facebook Prole Pictures. In Proceedings of the
ACM International Conference on Multimedia. ACM, 1101–1104.
[9]
Fabio Celli and Luca Rossi. 2012. e role of emotional stability in Twier
conversations. In Proceedings of the Workshop on Semantic Analysis in Social
Media. Association for Computational Linguistics, 10–17.
[10]
Paul T Costa and Robert R McCrae. 2008. e Revised NEO Personality Inventory
(NEO-PI-R). In G.J. Boyle, G Mahews and D. Saklofske (Eds.). e SAGE handbook
of personality theory and assessment 2 (2008), 179–198.
[11]
Jennifer Golbeck, Cristina Robles, Michon Edmondson, and Karen Turner. 2011.
Predicting personality from twier. In Privacy, Security, Risk and Trust (PASSAT)
and 2011 IEEE ird Inernational Conference on Social Computing (SocialCom),
2011 IEEE ird International Conference on. IEEE, 149–156.
[12]
Jennifer Golbeck, Cristina Robles, and Karen Turner. 2011. Predicting personality
with social media. In CHI’11 extended abstracts on human factors in computing
systems. ACM, 253–262.
[13]
Ashley S. Holland and Glenn I. Roisman. 2008. Big Five personality traits and
relationship quality: Self-reported, observational, and physiological evidence.
Journal of Social and Personal Relationships 25, 5 (2008), 811–829.
[14]
Wayne D. Hoyer and Steven P. Brown. 1990. Eects of brand awareness on
choice for a common, repeat purchase product. Journal of Consumer Research 17
(1990), 141–148.
[15]
Michal Kosinski, Yoram Bachrach, Pushmeet Kohli, David Stillwell, and ore
Graepel. 2014. Manifestations of user personality in website choice and behaviour
on online social networks. Machine learning 95, 3 (2014), 357–380.
[16]
Michal Kosinski, David Stillwell, and ore Graepel. 2013. Private traits and
aributes are predictable from digital records of human behavior. Proceedings of
the National Academy of Sciences 110, 15 (2013), 5802–5805.
[17]
Emma D. Macdonald and Byron M. Sharp. 2000. Brand awareness eects on
consumer decision making for a common, repeat purchase product: A replication.
Journal of Business Research 48 (2000), 5–15.
[18]
Wilhelm Marston. 1928. Emotions of normal people. Kegan Paul Trench, Trubner
& Co. Ltd.
[19]
Isabel Briggs Myers and Peter B Myers. 2010. Gis diering: Understanding
personality type. Davies-Black Publishing.
[20]
Andrew Neal, Gillian Yeo, Annee Koy, and Tania Xiao. 2012. Predicting the
form and direction of work role performance from the Big 5 model of personality
traits. Journal of Organizational Behavior 33, 2 (2012), 175–192. hps://doi.org/
10.1002/job.742
[21]
in Nguyen, Dinh Phung, Bre Adams, and Svetha Venkatesh. 2011. Towards
discovery of inuence and personality traits through social link prediction. In
Proceedings of the International AAAI Conference on Weblogs and Social Media.
566–569.
[22]
Sandeep Pandey, Mohamed Aly, Abraham Bagherjeiran, Andrew Hatch, Peter
Ciccolo, Adwait Ratnaparkhi, and Martin Zinkevich. 2011. Learning to target:
What works for behavioral targeting. In Proceedings of the 20th Conference on
Information and Knowledge Management (CIKM). ACM, 1805–1814.
[23]
Daniele ercia, Michal Kosinski, David Stillwell, and Jon Crowcro. 2011.
Our Twier proles, our selves: Predicting personality with Twier. In Privacy,
Security, Risk and Trust (PASSAT) and 2011 IEEE ird Inernational Conference on
Social Computing (SocialCom), 2011 IEEE ird International Conference on. IEEE,
180–185.
[24]
Brent W. Roberts, Nathan R. Kuncel, Rebecca Shiner, Avshalom Caspi, and
Lewis R. Goldberg. 2007. e Power of Personality: e Comparative Validity of
Personality Traits, Socioeconomic Status, and Cognitive Ability for Predicting
Important Life Outcomes. Perspectives on Psychological Science 2, 4 (2007), 313–
345. hps://doi.org/10.1111/j.1745-6916.2007.00047.x
[25]
Bakhtawar Seerat and Usman Qamar. 2015. Rule induction using enhanced
RIPPER algorithm for clinical decision support system. In Intelligent Control
and Information Processing (ICICIP), 2015 Sixth International Conference on. IEEE,
83–91.
[26]
Cristina Segalin, Fabio Celli, Luca Polonio, Michal Kosinski, David Stillwell, Nicu
Sebe, Marco Cristani, and Bruno Lepri. 2017. What your Facebook prole picture
reveals about your personality. In Proceedings of the International Conference on
ACM Multimedia (ACM-MM). ACM.
[27]
Cristina Segalin, Dong Seon Cheng, and Marco Cristani. 2017. Social Proling
rough Image Understanding. Comput. Vis. Image Underst. 156, C (March 2017),
34–50. hps://doi.org/10.1016/j.cviu.2016.10.013
[28]
Shirish Krishnaj Shevade, S Sathiya Keerthi, Chiranjib Bhaacharyya, and Kara-
turi Radha Krishna Murthy. 2000. Improvements to the SMO algorithm for SVM
regression. Neural Networks, IEEE Transactions on 11, 5 (2000), 1188–1193.
[29]
Jacopo Staiano, Bruno Lepri, Nadav Aharony, Fabio Pianesi, Nicu Sebe, and Alex
Pentland. 2012. Friends don’t lie: inferring personality traits from social network
structure. In Proceedings of the 2012 ACM conference on ubiquitous computing.
ACM, 321–330.
[30]
Simine Vazire and Samuel D Gosling. 2004. e-Perceptions: personality impres-
sions based on personal websites. Journal of Personality and Social Psychology
87, 1 (2004), 123.
[31]
Ju Yan, Ning Liu, Gang Wang, Wen Zhang, Yun Jiang, and Zheng Chen. 2009.
How much can behavioral targeting help online advertising?. In Proceedings of
the 18th International Conference on World Wide Web (WWW). ACM, 261–270.
[32]
Jun Yang, Yu G. Jiang, Alexander G. Hauptmann, and Chong W. Ngo. 2007.
Evaluating bag-of-visual-words representations in scene classication. In MIR ’07:
Proceedings of the international workshop on Workshop on multimedia information
retrieval. ACM, New York, NY, USA, 197–206. hps://doi.org/10.1145/1290082.
1290111
[33]
Wu Youyou, Michal Kosinski, and David Stillwell. 2015. Computer-based per-
sonality judgments are more accurate than those made by humans. Proceedings
of the National Academy of Sciences 112, 4 (2015), 1036–1040.
[34]
Xiaojin Zhu. 2006. Semi-supervised learning literature survey. Computer Science,
University of Wisconsin-Madison 2 (2006), 3.