Who’s Who with Big-Five: Analyzing and Classifying Personality
Traits with Smartphones
Gokul Chittaranjan
1,2
, Jan Blom
3
, Daniel Gatica-Perez
1,2
gthatta@idiap.ch, jan.blom@nokia.com, gatica@idiap.ch
1
Idiap Research Institute, Centre du Parc, PO Box 592, Rue Marconi 19, 1920 Martigny, Switzerland
2
Ecole Polytechnique Federale de Lausanne, 1015 Lausanne, Switzerland
3
Nokia Research Center Lausanne, PSE-C, EPFL, 1015 Lausanne, Switzerland
Abstract
In this paper, we investigate the relationship between
behavioral characteristics derived from rich smartphone
data and self-reported personality traits. Our data
stems from smartphones of a set of 83 individuals col-
lected over a continuous period of 8 months. From
the analysis, we show that aggregated features obtained
from smartphone usage data can be indicators of the
Big-Five personality traits. Additionally, we develop an
automatic method to infer the personality type of a user
based on cellphone usage using supervised learning.
We show that our method performs significantly above
chance and up to 75.9% accuracy. To our knowledge,
this constitutes the first study on the analysis and clas-
sification of personality traits using smartphone data.
1 Introduction
The rapid global growth of mobile phone usage [1]
has reinforced the need to study the psychological, so-
cial, and economic implications of mobile telephony.
Smartphones provide a new lens for investigating
mobile phone usage [22]. Such platforms are pro-
grammable, allowing the development of data collec-
tion tools to record various behavioral aspects of the
user, ranging from how the device is used across differ-
ent contexts to analyzing spatial and social dimensions
of the everyday life of the user, as captured through
sources like GPS, call logs, and Bluetooth. From the
point of view of designing communication features and
applications that are tailored to the individual needs
and preferences of a user, this data intensive framework
provides a wealth of new opportunities, as it allows us
to understand the impact of context on user behav-
ior as well as to study individual differences, such as
personality of the users.
In personality psychology, personality traits play a
central role in describing a person [17]. This topic has
also been found to be of vital importance in comput-
ing. Several studies have been recently conducted on
personality traits and their relationship to the use of
Internet and forms of social media such as Youtube,
blogs, Facebook and other social networks [3, 6, 24, 26].
Since mobile phones also mediate social interac-
tions, phone usage could reflect an individual’s person-
ality [5]. However, in contrast to the significant amount
of research in the web context, surprisingly few stud-
ies have been carried out in the past to investigate the
connection between mobile phone usage and personal-
ity of individuals. In particular, the following points
have not yet been adequately addressed: First, there
is a clear need for scalability of studies to both a large
and diverse feature set and a user base. This has not
been possible in the past because of the burden on a
user, who is often a customer, in answering lengthy
questionnaires. Second, the rich contextual informa-
tion that can be extracted with current smartphones
has not been studied from the perspective of personal-
ity. Third, the automatic inference of usage or traits,
based on features that can be reliably extracted from
continuously collected data has not been explored.
Determining the personality of mobile phone users,
besides being important solely from the psychological
point of view, can also provide an interesting frame-
work for wearable computing. The ability to draw con-
nections between behavioral aspects derived through
contextual data collected by mobile phones, as well
as personality, could lead to designing and applying
machine learning methods to classify users into per-
sonality types. Such understanding could be used in
various ways. For instance, prior research has shown
that personality is linked to user interface preferences,
1
like the surface color of an application [4]. Certain per-
sonality traits, like extraversion/introversion, have also
been found to be linked to preferences pertaining to vi-
sual aesthetics of web sites [11]. The personality of a
user might also determine the kind of functionality the
individual is disposed to use on the phone. Individ-
ual differences in personality may also correlate with
the impact of context on the user. For instance, when
faced with idle time, is an extraverted person likely to
use the device in a different way, as compared to an in-
trovert? The preferred interaction modalities may also
differ across personality types. Conscientious persons,
for example, may be prone to switch their devices to a
silent mode in a socially sensitive situation.
Although the examples given above are hypothet-
ical, they nevertheless indicate that expending efforts
on establishing a link between personality and behavior
can be justified by the wealth of design opportunities
such a discovery would lead to. In this work, we an-
alyze the relationship between smartphone usage and
the personality traits in the Big-Five Model [17], and
develop an automatic method to classify users accord-
ing to self-perceived personality using features that are
by nature privacy sensitive and extracted from usage
logs and phone sensors on the Nokia N95 smartphone.
Our experiments are based on subset of the Lausanne
Data Collection Campaign[12], and contains data con-
tinuously collected from 83 participants over 8 months
of time. First, we show that significant relationships
exist between personality traits and some of the ex-
tracted features using correlation and multiple regres-
sion analyses. Next, we describe an automated method
to classify users according to their personality traits.
We show that our method’s performance is promising.
To our knowledge, this constitutes the first study on
inferring personality traits of smartphone users from
automatically extracted features.
The paper is organized as follows. Section 2 de-
scribes past work relating to personality measurement
through direct or indirect means. The dataset used
along with details about feature extraction is given in
section 3. The correlation analysis between features
and personality, and subsequently the classification of
users based on their Big-Five traits is described in sec-
tion 4. Finally, we conclude in section 5.
2 Related Work
The Big-Five personality framework [17] has re-
ceived considerable support in psychology in the past,
although there has not been a universal acceptance of
the concept. This framework is a hierarchical model
of personality traits that represent personality at the
broadest level of abstraction [10]. It consists of five
bipolar factors, namely extraversion, agreeableness,
conscientiousness, neuroticism, and openness to expe-
rience [17]. These factors, described in Table 1, sum-
marize several more specific traits and are believed to
capture most of the individual differences in human
personality [10].
Given the objectives of this work, it is useful to
contrast personality assessment methods into ques-
tionnaire and behavior based. The questionnaires
used in many Big-Five personality studies are typically
lengthy. This can be a limitation when a large number
of participants at geographically spread areas have to
complete questionnaires online. Therefore, efforts have
been made to develop brief scales in psychology [10],
so as to minimize the time required by the participants
to fill in a survey as well as the cost associated with
the process of filling in questionnaires. In this context,
Gosling et. al. introduced the Ten Item Personality
Inventory (TIPI) [10] that includes, as the name sug-
gests, ten questions to determine the Big-Five personal-
ity traits. It has been shown that the TIPI instrument
reaches adequate convergence with the Big-Five mea-
sures in self-reported ratings [10]. Hence, in our study
we use TIPI to measure self-perceived personality.
On the other hand, in relation to assessing person-
ality indirectly, through behavioral characteristics Pi-
anesi et. al. showed that personality traits in a meeting
environment can be detected using audio-visual fea-
tures and supervised learning [19]. In this case, per-
sonality of the participants was revealed by how par-
ticipants spoke and interacted in the experimental sit-
uation. Similarly, Mairesse and Walker describe an au-
tomatic procedure using NLP and audio features to
detect the Big-Five traits from conversation extracts
[15, 16]. While the above examples highlight that be-
havior can be indicative of the personality of an indi-
vidual, the role of the mobile phone in revealing this be-
havior remains a relatively unexplored territory. This
is surprising given that there is plenty of prior research
pertaining to modeling users and their mobile phone
usage patterns. To name a few examples, Eagle and
Pentland described the concept of eigenbehavior and its
usefulness in predicting behavioral patterns and ties in
a network of people [8]. Farrahi and Gatica-Perez have
illustrated ways of determining routines of users by
modeling sensor data pertaining to location collected
from mobile phones using topic models [9]. Further,
Do and Gatica-Perez [7] recently presented an analysis
of application usage in smartphones, for the purpose
of user retrieval. Similarly, Verkasalo et. al. studied
the reasons and motivation behind using applications
across users and non-users [25]. These studies tie well
with the thriving “app-usage” culture established by
smartphone manufacturers - through services like the
Table 1: Big-Five traits and examples of adjectives [17]
Trait Examples of Adjectives
Extraversion (E) Active, Assertive, Energetic, Enthusiastic, Outgoing, Talkative
Agreeableness (A) Appreciative, Forgiving, Generous, Kind, Sympethetic
Conscientiousness (C) Efficient, Organized, Planful, Reliable, Responsible, Thorough
Neuroticism (N) Anxious, Self-pitying, Tense, Touchy, Unstable, Worrying
Openness to Experience (O) Artistic, Curious, Imaginative, Insightful, Original, Wide Interests
Apple App Store, Nokia Ovi Store and the Android
Market. However, very few studies have directly ad-
dressed the relationship between smartphone usage and
personality, although personality plays a vital role in
social science and psychology.
In a related study, Poschl and Doring presented an
analysis relating usage patterns in phones to users clus-
tered on the basis of big-five personality traits into
two discrete groups. All information in the study was
gathered using questionnaires [20]. Recently, Butt and
Phillips presented a study of personality and its rela-
tionship to mobile phone usage [5]. The detailed NEO-
FFI personality test in conjunction with the Cooper-
smith self-esteem inventory were administered to par-
ticipants of the study. Factors describing levels of
phone usage were obtained from another questionnaire.
The features used in this study were related to phone
calls and SMS usage. Many of the comparisons made
in the study were motivated by previous work inves-
tigating the link between personality traits and Inter-
net usage [5]. In this study, disagreeable individuals
tended to be more likely to report receiving more calls
and also a higher proportion of calls as “unwanted”.
Outgoing calls were not significantly explained by the
traits. Extraverted, neurotic, and non-conscientious in-
dividuals were reported to have spent more time send-
ing/receiving SMS, and extraverted and disagreeable
individuals were found to spend more time changing
the ring tone or wallpapers. In a similar work, Phillips
et. al. [18] also found that disagreeable individuals
were more likely to play games on their phone [18].
Our study differs from past work in several ways. We
utilize information available in today’s smartphones,
such as the usage of web, music, video, maps, proxim-
ity information derived from bluetooth etc., in addition
to the traditional call and SMS usage information. All
cues are automatically extracted from usage logs, with-
out intervention or input from users. Therefore, we do
not rely on personal recall of these usage cues, that can
be prone to human errors and biases. We use a short
personality questionnaire that makes the project scal-
able to a large population. Finally, we also device an
automatic method, using supervised learning to clas-
sify users according to the Big-Five traits.
3 Dataset and Feature Extraction
In our work, we use smartphone data of 83 partic-
ipants who participated in the Lausanne data collec-
tion campaign [12], a people sensing project organized
in French-speaking region of Switzerland. We use data
collected for a continuous period of 8 months (between
November 2009 and July 2010) using a continuous non-
intrusive, passive data collection software running on
Nokia N95 phones. This software collected anonymized
logs of calls (Call Logs), SMS (SMS Logs), bluetooth
scans (BT Logs), and application usage (App Logs).
Of the 83 participants, 53 were male and 28 were
female, 2 participants chose not to disclose their gen-
der. The mean age was 29.7 years with a standard
deviation of 7.6 years. The minimum and maximum
ages were 19 and 63 years respectively. 63 participants
had at least a university degree. Among the partici-
pants, 38 were Asians, 2 were North Americans and 42
were Europeans. All users were previous mobile phone
users, but most of them had not owned a smartphone
before the study. Therefore, they discovered most of
the features of the N95 phone during the study.
Self-perceived personality was measured using the
TIPI questionnaire [10]. This was obtained as a part
of an online exit survey administered in two languages,
based on the language preference. The TIPI question-
naire could be answered within a few minutes.
The features used in our studies are aggregated from
the logs on a monthly level. In other words, all users
were split across months, which gave us 567 user-
months. From each of the user-months, features de-
scribing different aspects of a smartphone usage were
computed automatically, by parsing the logs, as sum-
marized in Table 2. The selection of features is based
on previous work enlisted in section 2 and on the choice
of features that could reasonably characterize levels
of individual and relational activity. All features ex-
cept those from BT Logs were obtained by aggregat-
ing events (such as the opening of an Office or Inter-
net application) as and when they happened. Features
pertaining to Bluetooth, were based on scans done ap-
proximately once every 3 minutes. Defining a time slot
as one bluetooth scan, we also computed features that
captured the duration for which an ID was available.
Table 2: Usage cues extracted from smartphone data aggregated on a monthly basis
Source Usage cue
Application
No. of uses of Office Apps
No. of uses of Internet Apps
No. of uses of Video/Audio/Music Apps
No. of uses of Maps App
No. of uses of Mail App
No. of uses of Youtube App
No. of uses of Calendar App
No. of uses of Camera App
No. of uses of Chat Apps
No. of uses of SMS App
No. of uses of Games
Bluetooth(BT)
Number of unique BT IDs
Times most common BT ID is seen
BT IDs accounting for 50 % of IDs seen
†
BT IDs seen for more than 4 time slots
†
BT IDs seen for more than 9 time slots
†
BT IDs seen for more than 19 time slots
†
Max. time for which a BT ID is seen
†
SMS
Average word length (Inbox)
Median word length (Inbox)
No. messages (Inbox)
†
See Section 3 for description
Source Usage cue
Calls
No. of outgoing (O) calls
Average duration of outgoing calls
Total duration of outgoing calls
No. of incoming (I) calls
Average duration of incoming calls
Total duration of incoming calls
No. of unique contacts called
No. of unique contacts who called
No. of incoming and outgoing calls
Average duration of I and O calls
Total duration of I and O calls
No. of unique contacts communicated with
No. of missed calls
No. of contacts associated with missed calls
No. of SMS received
No. of unique contacts SMS received from
No. of SMS sent
No. of unique contacts SMS sent to
SMS
Average word length (Sent)
Median word length (Sent)
No. messages (Sent)
4 Experiments and Results
We systematically analyze the relationship between
personality traits and usage features in the following
section, using standard correlation and multiple regres-
sion analyses. Next, we present a machine learning ap-
proach to classify users according to their big-five traits
using SVM and C4.5 classifiers.
4.1 Data Analysis
4.1.1 Analysis of independent variables
The mean and standard deviation for the five traits,
along with the minimum and maximum value and
skewness is tabulated in Table 3. It should be noted
that Neuroticism is replaced by Emotional Stability,
since this was computed from the TIPI questionnaire.
None of the predictors had a skew ≥ 1 or ≤ −1, hence
no transformation was applied to the data.
The correlations between the traits are given in Ta-
ble 4. Although significant correlations (as high as 0.64
between agreeableness and emotional stability) exist,
all were below the selection criteria used in the test for
multi-collinearity in previous work [5].
4.1.2 Analysis of Dependent Variables
All the automatically extracted features were found to
be positively skewed, so a log-transformed transforma-
tion, log(feature + 1) was applied for the correlation
and multiple regression analyses. We first found that
the features derived from App Logs were very sparse,
due to the low usage frequency of some applications.
Hence, for analyses involving App Logs, we chose only
those user-months for which there has been some use
of a given application. Further, for all analyses, only
those user-months with at least 7 days of usage of ei-
ther calls, SMS, BT or applications were chosen. This
was done to avoid user months that might contain no
data, owing to a variety of reasons such as vacations,
problems with phone use etc.
Pearson’s correlation coefficient between the fea-
tures and the traits, as well as multiple regression anal-
yses between each of the features against all the inde-
pendent variables (traits) were done. These results are
tabulated in Tables 5 and 6 respectively. The standard-
ized regression coefficient (β) and t−statistics were also
computed. For purposes of clarity, the discussion to fol-
low has been organized based on the source of features.
App Logs: It was found that the usage of all appli-
cations, except the use of Maps, Camera, Chat and
Game applications significantly explained variance in
the traits. Multiple regression analysis showed that
the traits accounted for 12% of the variance in the use
of Office, with conscientious (β = 0.25, t = 3.01, p =
0.003), not emotionally stable (neurotic) (β = −0.22,
t = −2.29, p = 0.023) and low-openness (β = −0.29,
t = −3.97, p < 0.001) participants to use it more.
Table 3: Statistics for the Big-Five traits
Predictors Mean SD Min. Max. Skew
Extraversion 4.09 1.24 1.00 7.00 -0.18
Agreeableness 4.88 1.28 2.00 7.00 -0.39
Conscientiousness 5.10 1.42 1.00 7.00 -0.71
Emotional Stability 4.61 1.24 1.50 7.00 -0.28
Openness to Experience 4.69 1.42 1.00 7.00 -0.64
Table 4: Correlations across the 5-independent variables (Big-Five Traits)
Predictors 1 2 3 4 5
1. Extraversion -0.06 -0.07 -0.29
∗∗
0.11
2. Agreeableness 0.50
∗∗
0.64
∗∗
0.41
∗∗
3. Conscientiousness 0.50
∗∗
0.50
∗∗
4. Emotional Stability 0.39
∗∗
5. Openness to Experience
∗∗
p < 0.01
Further, we found that the Mail application was more
likely to be used by neurotic (β = −0.27, t = −2.32,
p = 0.021) and conscientious (β = 0.33, t = 2.92,
p = 0.004) participants. This trend is similar to that
observed with the use of office applications. A poten-
tial explanation to this correlation is that these appli-
cations are used with relation to work. Hence partici-
pants using e-mail to communicate might be expected
to be more responsible and efficient. Interestingly, in-
troverts (β = −0.2575, t = −5.246, p < 0.001) were
less likely to use Internet applications on the phone.
This is also reinforced by the significant negative pair-
wise correlation observed in Table 5 (r = −0.26, p <
0.01). Although the use of Audio/Video/Music appli-
cations was explained only to an extent of 4% by the
traits, upon examining the regression coefficients, we
observed that conscientious individuals were less likely
to use them (β = −0.21, t = −3.735, p < 0.001). This
relationship also shows up in the form of a significant
negative correlation in Table 5 (r = −0.18, p < 0.01).
Finally, the SMS application (which might not cor-
respond to the actual number of SMS sent/received)
was found to be more likely to be used by disagreeable
(β = −0.15, t = −2.61, p = 0.01), conscientious partic-
ipants (β = 0.16, t = 3.07, p = 0.002) who score low on
openness (β = −0.24, t = −4.74, p < 0.001). Consci-
entious individuals are planful, thorough and responsi-
ble, while disagreeableness relates to being unforgiving
and showing low trust. These factors could have con-
tributed to the SMS applications usage trends.
BT Logs: None of the BT features explained a large
percentage of variance in the traits. We observed that
introverts (β = −0.12, t = −2.60, p = 0.010) and
disagreeable (β = −0.15, t = −2.48, p = 0.01) partic-
ipants had fewer number of unique BT IDs scanned.
This could be because of introverts and disagreeable
individuals being less social in nature. The number of
BT IDs seen multiple times (4, 9, and 19) showed a
consistent trend with positive β values for emotional
stability and negative β values for agreeableness, sug-
gesting that emotionally stable and disagreeable indi-
viduals might encounter BT IDs for longer durations.
This might be due to emotionally stable and disagree-
able participants having a tendency of staying in the
same place for longer periods of time.
SMS Logs: Multiple regression analysis on our data
showed that message size in the inbox were not sig-
nificantly explained by the traits. However, the num-
ber of messages in the inbox showed that emotionally
stable users (β = 0.18, t = 2.86, p = 0.004) scoring
low on openness (β = −0.15, t = −3.00, p = 0.002)
were more likely to have more messages. Features
describing the quality of messages in the sent folder
had their variance explained significantly by the traits.
Emotionally stable individuals (For Avg. word length
feature, β = 0.16, t = 2.61, p = 0.009) with low-
openness (For Avg. word length feature, β = −0.19,
t = −3.77,p < 0.001) were found to use longer words.
They also had more items in the sent folder. This sug-
gests that the personality of a user may be better un-
derstood using SMS composed by him, rather than the
messages he receives.
Call Logs: While the number and duration of outgo-
ing calls was not significantly explained by the five
traits, the variance in the number and duration of in-
coming calls and the number of unique contacts as-
sociated with them, were significantly explained. Ex-
traverts (β = 0.18, t = 3.98, p < 0.001) and agree-
able (β = 0.15, t = 2.59, p = 0.01) individuals were
likely to receive more calls. This is also seen in pair-
wise correlations with r = 0.13 and 0.20 for extraver-
sion and agreeableness respectively. The duration of
![Table 1: Big-Five traits and examples of adjectives [17]](/figures/table-1-big-five-traits-and-examples-of-adjectives-17-2j97xs7a.webp)
