User Personality and the New User Problem in a Context-Aware Point of Interest Recommender System

Full text

User Personality and the New User Problem in a Context-
Aware Points of Interest Recommender System
Matthias Braunhofer,
Mehdi Elahi, and
Francesco Ricci
Faculty of Computer Science
Free University of Bozen - Bolzano, Italy
mbraunhofer@unibz.it, mehdi.elahi@unibz.it, fricci@unibz.it
Abstract
The new user problem is an important and challenging issue that Context-Aware Recommender
Systems (CARSs) must deal with, especially in the early stage of their deployment. It occurs
when a new user is added to the system and there is not enough information about the user’s
preferences in order to compute appropriate recommendations. It is common to address this
problem in the recommendation algorithm, by using demographic attributes such as age,
gender, and occupation, which are easy to collect and are reasonably good predictors of the user
preferences. However, as we show here, user’s personality provides even better information for
generating context-aware recommendations for places of interest (POI), and it is still easy to
assess with a simple questionnaire. In our study, using a rating data set collected by a mobile
app called STS (South Tyrol Suggests), we have found that by considering the user personality
the system can better rank the recommendations for the new users.
Keywords: Context-Aware Recommender Systems; Cold-Start Problem; New User Problem;
Personality; Demographics
1 Introduction
With the tremendous growth of information available on the World Wide Web, it has
become more and more difficult to search for and find relevant information on the
Web. For instance, users of online tourism portals often find it difficult to select a
hotel or a place to stay because of the overwhelming number of travel offers and the
lack of effective system support. Recommender Systems (RSs) help overcome this
information overload by providing users with selected (information) items that match
their personal preferences (Ricci et al., 2011). The suggested items can be obtained by
comparing the user’s profile (which appropriately models the user preferences) with
the descriptions of items (content-based approach) or with the profiles of other users
(collaborative-filtering approach).
Context-Aware Recommender Systems (CARSs) are a special type of RSs that
generate more relevant recommendations by adapting them to the specific contextual
situation of the user (e.g., weather, temperature, season, and companion). Over the
last couple of years, increasingly more elaborate recommendation algorithms and
techniques for incorporating contextual information into the recommendation process
have been proposed. According to (Adomavicius et al., 2011), they can be classified
as contextual pre-filtering (or contextualisation of the recommendation input),
contextual post-filtering (or contextualisation of the recommendation output) and
contextual modelling (or contextualisation of the recommendation function),

depending on the computational stage where the available contextual information is
used.
Nevertheless, the cold-start problem remains a demanding challenge for CARSs as
well. This problem occurs when: 1) a new user without rating history requests a
recommendation from the system (new user problem); 2) the RS is asked to generate
a recommendation for an unrated item (new item problem); or 3) the RS is requested
for recommendations under contextual situations that have not been rated before (new
context problem).
This paper focuses on the new user problem, which from the user’s point of view is
the most important one as it prevents new users from obtaining any recommendation
tailored to their interests and needs. A common approach to this problem is either to
provide non-personalised recommendations (e.g. popular items) or to utilize simpler
recommendation techniques, for instance, those based on demographic user data, such
as gender, age group, Zip code and income level. For instance, based on the
knowledge that “males like soccer” and “females like figure skating”, information
items related to soccer and figure skating can be suggested to users with those
demographics, respectively.
In this paper, we propose to consider the users’ personality to overcome the new user
problem in CARSs. We model the personality with the five-factor model (FFM) or
Big-5 personality traits (Costa & McCrae, 1996), which is currently the most
widespread and widely accepted approach. According to this model, user personality
is measured along five dimensions: openness, conscientiousness, extraversion,
agreeableness, and neuroticism. The main reason for using personality in RSs is that it
is a predictable and stable factor that explains human behaviours. It has been shown
that there exist direct links between personality and interests / preferences (Rentfrow
et al., 2003), and hence, people with similar personality traits have similar interests /
preferences. Moreover, user personality can be acquired through self-report
questionnaires, such as the Five-Item Personality Inventory (FIPI) (Gosling et al.,
2003), which we have used, or the 120 or 240 item International Personality Item
Pool Representation of the NEO PI-R (IPIP-NEO) (Goldberg et al., 2006).
In a previous paper (Elahi et al., 2013) we have used the personality not only to
generate better recommendations, but also to identify which ratings to acquire from
the user, i.e., which POIs are useful for users to rate because they better describe the
user preferences. The proposed rating elicitation technique leads to a significant
increase in the number of ratings obtained from the users as well as the improvement
of the recommendation accuracy. In this paper we further develop that analysis by
comparing the benefit of exploiting the user personality to the usage of demographic
information of the user, which is a more common approach. We investigate the effect
of each of the five considered personality factors separately in order to determine
which of them largely contribute to the accuracy of the recommendation model. To
the best of knowledge, this is a novel work especially in the tourism domain.
In summary, in this article we present the evaluation of the usage of the Big-5
personality traits (Costa & McCrae, 1996) in a state-of-the-art CARS algorithm, i.e.,
CAMF-CC (Baltrunas et al., 2011), and we compare it to a similar model that uses
demographic attributes (i.e., age group and gender). Experiments were executed using

the rating dataset derived from the South Tyrol Suggests (STS) app (Braunhofer et al.,
2014a), an Android-based CARS that provides users with place of interest (POI)
recommendations using various contextual factors (e.g., weather, time of day,
location, and mood). The results show that the ranking quality and the prediction
accuracy of CARS can be improved by exploiting personality characteristics that cut
across demographic categories.
The rest of this paper is organized as follows. Section 2 reviews the state of the art. In
Section 3, we explain the implemented procedure to collect the demographics and
user personality in our STS app. In Section 4, we present the experimental evaluation
and the obtained results. Finally, in Section 5 we elaborate conclusions and give
future work directions.
2 Related Work
Research on CARSs is a rather fresh topic that has drawn considerable attention in the
last years. These systems recommend items personalised for each user by considering
the contextual situation of the item and the user (Adomavicius et al., 2011). For
instance, liveCities (Martin et al., 2011) is a CARS that, by considering the location of
the tourists, sends them push-based notifications such as a suggestions about what to
do, or discounts offered by nearby restaurants. In fact, various CARSs exploits the
contextual information of the tourists to provide them with more relevant suggestions
(Costa et al., 2013; Meehan et al., 2013; Barranco et al., 2012).
While recommendation techniques exploited in CARSs can be effective and robust
for their frequent users, all of them have poor performance for new users and new
items (cold-start problem). Particularly, in collaborative filtering based recommender
systems, which are the type of recommenders that we are considering in this work, the
cold-start problem is determined by the lack of a sufficient number of ratings for the
system to be able to compute relevant and diverse suggestions.
This problem is even worse for CARSs. This is because, in order to properly function,
these systems must collect not only ratings given by users to items, but also the
description of the contextual situation under which the items were experienced by the
users and that can have influenced the users appreciation of the items. For instance,
consider a CARS that has obtained many low ratings for a mountain hiking route
when the weather context of the user experience of the route was “rainy” (which here
we assume that influenced negatively the ratings). However, suppose that the system
has never collected any rating when the weather context was “sunny”, which is a
situation that we assume here can be more suited for that experience. In such a case,
that route would not be recommended to the users in a sunny day since the system has
no knowledge of how the users rate that route in a sunny day and it is not able to
make a correct extrapolation from the low ratings given in the “rainy” contextual
situation.
A common solution of the cold-start problem in RSs, relies on the usage of
demographic information of the users. For example, (Pazzani, 1999) computed the
user-to-user similarity in a neighbor-based collaborative filtering system by using
demographic attributes of the users, e.g., gender, age group, area code, education, or

employment. This is an example of a more general class of techniques that is
sometimes called Demographic Filtering (Adomavicius et al, 2005), which is often
used in tourism applications. For example, (Wang et al, 2012) applies demographic
filtering to categorize the tourists by their demographic attributes and generate
recommendations based on their demographic classes. The main advantage of such
systems is that they do not require any rating or extra knowledge from the users to
make recommendations for them.
A more novel line of solutions of the cold-start problem is based on the exploitation
of personality information of the users. As mentioned before, personality is a user
characteristic that influences very much decision making as well as tastes and
interests (Rentfrow et al., 2003; John, 1999). In fact, people with similar personality
usually share similar interests and tastes.
In order to be able to incorporate personality information of the users in the
recommendation process, the system must identify and quantify the personality of the
user. This can be done either explicitly, by asking the user to complete a personality
questionnaire, or implicitly, by observing users’ behavioral patterns during the
interaction with the system (Kosinski et al., 2013). It has been shown that explicit
personality acquisition has higher accuracy and yields to better results in terms of
various measures such as user satisfaction, ease of use, and prediction accuracy (Dunn
et al., 2009). Moreover, explicit acquisition of user personality is easy. In (Hu et al.,
2009) it has been shown that the users spent significantly less effort, in terms of
perceived cognitive or actual task easiness and time, to complete the preference
elicitation stage in a personality based recommender system than in a rating-based
recommender system. In other words, it is also easier for the user to reply to questions
aimed at scoring the users in a personality model rather than evaluating items in given
rating scale (e.g., five star). Users have also expressed a stronger intention for re-
using the personality-based system. Hence, using personality characteristics of the
user can be a valuable source of information for the recommender systems for better
recommendation generation (Hu et al., 2011; Hu et al., 2009; Tkalcic et al., 2013).
3 South Tyrol Suggests (STS)
South Tyrol Suggests (STS1) is an Android-based mobile application that provides
tourists with recommendations for approximately 27,000 POIs (i.e., accommodations,
attractions, events, public services and restaurants) in South Tyrol (Italy) by
exploiting various contextual factors (e.g., weather, time of day, day of week,
location, and mood) in a recommendation algorithm that is based on matrix
factorisation (Koren & Bell, 2011). In particular, in STS, we have extended the
CAMF-CC model proposed by Baltrunas et al. (2011). This is a context-aware matrix
factorisation approach that integrates additional parameters for each contextual
condition and item category combination, besides the standard parameters used in
context-free matrix factorisation. We enhanced the original CAMF-CC model by
incorporating also demographics and/or personality information of the users, which
allows gathering behavioural information regardless of the availability of explicit
1 South Tyrol Suggests (STS): https://play.google.com/store/apps/details?id=it.unibz.sts.android

ratings from the user. This approach follows an analogous solution described in
(Koren & Volinsky, 2009). Given a user u, an item i, a contextual situation described
by the contextual conditions c1,…,ck and the set of attributes T(i) associated to item i,
it predicts ratings using the following formula:
𝑟
!"!!…!!=𝚤+𝑏!+𝑏!!!
!
!!!!∈!(!)
+𝑞!
!𝑝!+𝑦!
!∈!!
,
(1)
where qi, pu and ya are the latent factor vectors representing the item i, the user u and
the user attribute a (i.e., a particular gender, age group or Big-5 personality trait
score), respectively. ī is the average rating for item i, bu is the bias associated to user
u, and btcj is a parameter that models how the contextual condition cj influences the
rating of items belonging to category t. It takes a positive (negative) value depending
on whether the contextual condition cj has a positive (negative) effect on the ratings
for items belonging to category t. This allows, for instance, recommending indoor
POIs (e.g., museums, churches, castles) on bad weather conditions and outdoor POIs
(e.g., lakes, mountain hikes, scenic walks) on good weather conditions.
The model parameters are learned offline, once every five minutes, by minimising the
associated regularised squared error function through stochastic gradient descent
(Koren & Bell, 2011). This learning procedure is fast; it takes only a few seconds to
re-train the prediction model when new ratings are acquired.
3.1 Demographic Information
During the registration process STS asks to its users the following demographic data
(as can be seen in Figure 1, left): gender and birthday. Gender is given in a nominal
scale (female/male) and can be easily integrated into the extended matrix factorization
model of STS (as shown in Equation 1), in order to enhance the user representation.
Birthday, on the other hand, is an ordinal variable and we converted it to a less fine-
grained scale, just using 6 levels: <18, 18-25, 26-39, 40-54, 55-69, and 70+. These
age groups can then be used as additional user attributes integrated in the rating
prediction factor model, analogously to the gender attribute.
To better illustrate this, let us assume that the prediction model is asked to predict the
rating a female user u in the 18-25 age group would give to the item i, given that the
contextual situation c1,…,ck holds. In this case, the rating prediction formula is
rewritten as follows:
𝑟
!"!!…!!=𝚤+𝑏!+𝑏!!!
!
!!!!∈!(!)
+𝑞!
!𝑝!+𝑦!"#$%" +𝑦!"!!" .
(2)
Here, user u is now characterized by 3 components: the latent factor vector pu (which
is learned based on her prior ratings, if any), the latent factor vector yfemale (which is
learned from prior ratings given by female users) and the latent factor vector y18-25
(which is learned from prior ratings of users that are between 18 and 25 years old).

3.2 Personality Information
As part of the registration process, users are also asked to fill out the Five-Item
Personality Inventory (FIPI) (Gosling et al., 2003) so that the system can assess their
Big-5 personality traits, i.e., openness to experience, conscientiousness, extraversion,
agreeableness and neuroticism. Figure 1 (middle) shows a screenshot of STS where
one of the questionnaire statements (i.e., “I see myself as open to experience,
imaginative.”) is displayed. The full FIPI questionnaire takes approximately one
minute to complete, and consists of the following five statements to be rated on a 7-
point Likert scale from “strongly disagree” to “strongly agree”:
1. I see myself as open to experience, imaginative;
2. I see myself as dependable, organized;
3. I see myself as extraverted, enthusiastic;
4. I see myself as agreeable, kind;
5. I see myself as emotionally stable, calm.
The score for a specific personality trait is calculated by dividing the score (e.g., 0 for
“strongly disagree”, 1 for “disagree moderately”, …, 6 for “strongly agree”) by 6 (i.e.,
the maximum score), which gives a result in the interval [0, 1]. This result is then
discretized into 5 categories (i.e., “very high”, “high”, “neutral”, “low”, “very low”)
so that it can be integrated into the rating prediction model of STS, in order enhance
the user representation as well.
To see how this works, let us again consider an imaginary user u that scored high on
extraversion (= ext) and low on openness to experience (= ope), conscientiousness (=
Fig. 1. Sample screenshots of STS

con), agreeableness (= agr) and neuroticism (= neu). Then, her rating prediction
formula for item i and contextual situation c1,…,ck becomes as follows:
𝑟
!"!!…!!=𝚤+𝑏!+𝑏!!!+
!
!!!!∈!(!)
𝑞!
!𝑝!+𝑦!"#_!"# +𝑦!"#_!"# +𝑦!"#_!!"!+𝑦!"#_!"# +𝑦!"#_!"# .
(3)
Now, user u is profiled by 6 components: the latent factor vector pu and the latent
factor vectors yope_low, ycon_low, yext_high, yagr_low and yneu_low, which are derived from the
prior ratings of users that scored high on extraversion and low on openness,
conscientiousness, agreeableness and neuroticism, respectively.
4 Evaluation Methodology
In order to compare the effectiveness of the context-aware recommendation model of
STS when it is using the demographic information vs. the personality information, we
performed an offline experiment. The goal was two-fold: (1) to compare the
recommendation accuracy for new users that is achieved by using the demographics
information vs. that achieved by using the Big-5 personality traits; and (2) to identify
the demographic attributes or the Big-5 personality traits that, if introduced in the
prediction model, cause it to generate more accurate recommendations for new users.
4.1 Dataset
The rating dataset that we have used in the experiments was collected in about one
year of activity of our STS app. In total, the STS dataset contains 2,534 ratings
expressed on an ordinal scale from 1 (= “I didn’t like it”) to 5 (= “I liked it very
much”), with the steps of 1. Many ratings (i.e., 64.17 %) contained in this dataset are
augmented with the information about the contextual situations (at the time the items
were experienced) described by up to 14 contextual conditions (e.g., weather,
temperature, season, daytime, companion), which could be specified by the user by
means of an appropriate graphical user interface within the application, as shown in
Figure 1 (right). Moreover, the STS dataset contains basic demographics (i.e., age
group and gender) and personality information (i.e., Big-5 personality trait scores) of
the users.
Some of the ratings contained in the STS dataset were entered by users for which no
demographic and personality information was available. This happened when users
did not enter birthday/gender data on the registration screen or skipped the personality
questionnaire. Since in the evaluation, we wanted to compare the effect of using
personality information vs. using demographics in the CAMF model, we considered
only ratings from users for which both demographics and personality information
were available. In this way we obtained a dataset of 1,379 ratings, whose details are
summarized in Table 1.

Table 1. STS dataset information
Total number of ratings
1,379
Number of users
239
Number of items
184
Number of contextual factors
14
Number of contextual conditions
56
Number of contextual situations
799
Number of demographic attributes
2
Number of personality attributes
5
4.2 Evaluation Procedure
In order to compare the effectiveness of the system to produce recommendations for
new users in the two mentioned conditions, we carried out a ten-fold cross-validation
scheme as in (Braunhofer et al., 2014b, Shani et al., 2008). First, we randomly split
the users in the entire rating dataset into ten mutually exclusive subsets of
approximately equal size. Then, in each cross-validation run, we used the ratings
coming from one of the obtained user splits as testing set and the remaining ones as
training set to train the several distinct variants of the CAMF-CC model. In this
manner, we created a test set of ratings coming from users that have no ratings in the
training set, i.e., really cold (new) users without any observed rating. For each of the
ten iterations, two performance metrics were recorded:
1. Mean Absolute Error (MAE) (Shani & Gunawardana, 2011): This metric
evaluates the accuracy of predicted ratings. This is achieved by predicting
the rating 𝑟
!"!!…!! for each user-item-context triple (u, i, c1…ck) in the test set
T and comparing it with the actual rating 𝑟
!"!!…!!, as it is done in the
following formula:
𝑀𝐴𝐸 =
1
|𝑇||𝑟
!"!!…!!−𝑟
!"!!…!!|
(!,!,!!…!!)∈!
(4)
2. Normalized Discounted Cumulative Gain (nDCG) (Shani & Gunawardana,
2011): nDCG is a measure used to determine the effectiveness of RSs in
correctly ranking the recommendations compared with an optimal ranking.
Both, nDCG and MAE, were calculated based on the commonly accepted test
procedure proposed by Park & Chu (2009), which attempts to avoid performance bias
that is caused by the overrepresentation of high raters. It does so by clustering items
based on the ratings given by each user, and then randomly sampling one item for
each cluster to be used as test item. MAE and nDCG1 were then calculated on the
basis of these test items. We also calculated nDCGk, with k = 2, . . . 5, but for lack of
space and since they gave similar results we omit them.

Finally, the obtained MAE and nDCG estimates were averaged over all 10 runs, to
produce cross-validation estimates of MAE and nDCG.
4.3 Evaluation Results
The obtained results are shown in Table 2. As can be noted, profiling new users
through personality information led to more accurate recommendations, compared to
profiling them by demographics. Overall, the context-aware rating prediction model
achieved an MAE of 0.968 (SD = 0.079) by using the users’ Big-5 personality trait
information, whereas it achieved an MAE of 0.970 (SD = 0.081) by relying on
demographics (the lower the better). Similar results were obtained in terms of nDCG;
nDCG was higher by using personality than by using demographics (M = 0.723, SD =
0.055 vs. M = 0.665, SD = 0.065) (the higher the better). Subsequent t-tests performed
on both MAE and nDCG differences between models revealed that the MAE
difference was not significant (p = 0.452), whereas it was statistically significant the
difference of nDCG (p = 0.001). We believe that with a larger rating dataset also the
difference in MAE will likely become statistically significant. Hence, these
observations support our hypothesis that recommendations for new users are more
effective when they are generated based on the user’s personality type.
Table 2. Evaluation results
Type
Attribute
MAE
nDCG
Total MAE
Total nDCG
Demographics
Gender
0.969
0.679
0.970
0.665
Age group
0.975
0.676
Personality
Extraversion
0.968
0.680
0.968
0.723
Agreeableness
0.971
0.671
Conscientiousness
0.973
0.694
Emotional stability
0.959
0.704
Openness to experience
0.979
0.689
Moreover, Table 2 shows a comparison of the MAE and nDCG obtained by
enhancing the user representation in the context-aware rating prediction model with
individual demographic attributes (i.e., gender and age group) and personality traits
(i.e., extraversion, agreeableness, conscientiousness, emotional stability and openness
to experience). It can be seen that the best recommendation performance in terms of
both MAE (M = 0.959, SD = 0.092) and nDCG (M = 0.704, SD = 0.046) is achieved
using the emotional stability factor. This shows that emotional stability can be a
significant predictor for the user’s POI preferences, which is in line with previous
findings that showed emotional stability to be a good predictor of media preferences
(Hertel et al., 2008) and music preferences (Rentfrow & Gosling, 2006).

To sum up, the obtained results seem encouraging as they suggest that adapting the
recommendations to the target user’s personality profile can increase the system
recommendation effectiveness.
5 Discussions and Future Works
In this paper we have illustrated the benefits of incorporating the user personality
information in order to tackle the cold-start problem in a mobile context aware
recommender system for tourism called South Tyrol Suggests. We have analysed the
impact of utilizing personality traits, either individually or together, on the quality of
the recommendations, measured with well-known quality metrics, i.e., MAE
(indicator of rating prediction error) and nDCG (indicator of ranking quality). We
have shown that utilizing personality information is more effective than utilizing
demographic information of users, which is a more common approach to tackle the
cold-start problem. Moreover, we have shown that utilizing even a single trait out of
the five personality traits can still result in a significant improvement of the
recommendation quality. This last result implies that the personality traits are not
equally informative of the users long-term behavioural characteristics. Hence,
identifying the more important traits, that better represent and model the users’
characteristics, can be beneficial in designing and implementing better and easier to
apply personality based recommendation algorithms.
There are several research directions that we can consider for future work. First, we
are interested in analysing the usage of mobile phones and smart gadgets, such as
smart bracelets, in order to predict the user personality. In fact, it has been shown that
there are relations between people’s personality traits and the usage behaviours they
exhibit, which can be monitored by the above-mentioned devices. For example,
(McAffee et al., 2011) showed that extravert people access more the texting
functionality of the smartphones while users with higher agreeableness exploit more
the calling functionality. Therefore, using the usage data collected by smart gadgets
(e.g., how, when, where the gadget is used), one can predict the user personality and
then, as shown in this article, generate more relevant recommendations.
Moreover, we would like to repeat our offline experiments by using other datasets,
such as that collected by MyPersonality, a Facebook app for taking psychometric tests
by users. The users who took the tests using this app came from different age groups,
backgrounds and cultures. That makes the dataset unique and valuable. This dataset,
currently, has more than 6,000,000 test results together with more than 4,000,000
Facebook profiles of the users who took the tests. (My Personality Project, 2014). For
instance, there are more than 900,000 results for the big-five personality test. By
repeating our experiments on this larger dataset we can obtain a stronger confirmation
of our working hypotheses.
Finally, we would like to investigate the possibility of mining the social network
accounts of the users of STS in order to identify their personality. In fact, it has been
shown that the personality of the users can be learnt from their interactions in social
networks (Bachrach et al., 2012), and afterward, be used in collaborative filtering
based RSs (Fernández-Tobías et al., 2014). This is particularly interesting for the
users that skip the personality questionnaire during the registration phase.

References
Adomavicius, G., Mobasher, B., Ricci, F., & Tuzhilin, A. (2011). Context-Aware
Recommender Systems. AI Magazine, 32(3), 67-80.
Adomavicius, G., & Tuzhilin, A. (2005). Toward the next generation of recommender systems:
A survey of the state-of-the-art and possible extensions. Knowledge and Data
Engineering, IEEE Transactions on, 17(6), 734-749.
Baltrunas, L., Ludwig, B., & Ricci, F. (2011). Matrix factorization techniques for context aware
recommendation. In Proceedings of the fifth ACM conference on Recommender
systems (pp. 301-304). ACM.
Bachrach, Y., Kosinski, M., Graepel, T., Kohli, P., & Stillwell, D. (2012). Personality and
patterns of Facebook usage. In proceedings of the 3rd annual ACM web science
conference (pp. 24-32). ACM.
Barranco, M. J., Noguera, J. M., Castro, J., & Martínez, L. (2012). A context-aware mobile
recommender system based on location and trajectory. In Management intelligent
systems (pp. 153-162). Springer Berlin Heidelberg.
Braunhofer, M., Elahi, M., Ricci, F., & Schievenin, T. (2014a). Context-aware points of interest
suggestion with dynamic weather data management. In Information and
Communication Technologies in Tourism 2014 (pp. 87-100). Springer International
Publishing.
Braunhofer, M, Codina, V., & Ricci, F. (2014b). Switching hybrid for cold-starting context-
aware recommender systems. In Proceedings of the eigth ACM conference on
Recommender systems. ACM.
Costa, H., Furtado, B., Pires, D., Macedo, L., & Cardoso, A. (2013). Recommending POIs
based on the User’s Context and Intentions. In Highlights on Practical Applications of
Agents and Multi-Agent Systems (pp. 166-177). Springer Berlin Heidelberg.
Costa, P., & McCrae, R. (1996). Toward a new generation of personality theories: Theoretical
contexts for the five-factor model. The five factor model of personality: Theoretical
perspectives. Hrsg.: JS Wiggins. New York, 51-87.
Dunn, G., Wiersema, J., Ham, J., & Aroyo, L. (2009). Evaluating interface variants on
personality acquisition for recommender systems. In User Modeling, Adaptation, and
Personalization (pp. 259-270). Springer Berlin Heidelberg.
Elahi, M., Braunhofer, M., Ricci, F., & Tkalcic, M. (2013). Personality-based active learning
for collaborative filtering recommender systems. In AI* IA 2013: Advances in Artificial
Intelligence (pp. 360-371). Springer International Publishing.
Elahi, M., Ricci, F., & Rubens, N. (2014). Active Learning in Collaborative Filtering
Recommender Systems. In E-Commerce and Web Technologies (pp. 113-124).
Springer International Publishing.
Fernández-Tobías, I., & Cantador, I. (2014). Personality-Aware Collaborative Filtering: An
Empirical Study in Multiple Domains with Facebook Data. In E-Commerce and Web
Technologies (pp. 125-137). Springer International Publishing.
Goldberg, L. R., Johnson, J. A., Eber, H. W., Hogan, R., Ashton, M. C., Cloninger, C. R., &
Gough, H. G. (2006). The international personality item pool and the future of public-
domain personality measures. Journal of Research in Personality, 40(1), 84-96.
Gosling, S. D., Rentfrow, P. J., & Swann Jr, W. B. (2003). A very brief measure of the Big-
Five personality domains. Journal of Research in personality, 37(6), 504-528.
Hertel, G., Schroer, J., Batinic, B., & Naumann, S. (2008). Do shy people prefer to send e-
mail?: Personality effects on communication media preferences in threatening and
nonthreatening situations. Social Psychology, 39(4), 231.
Hu, R., & Pu, P. (2011). Enhancing collaborative filtering systems with personality
information. In Proceedings of the fifth ACM conference on Recommender systems
(pp. 197-204). ACM.

Hu, R., & Pu, P. (2009). A comparative user study on rating vs. personality quiz based
preference elicitation methods. In Proceedings of the 14th international conference on
Intelligent user interfaces (pp. 367-372). ACM.
John, O. P., & Srivastava, S. (1999). The Big Five trait taxonomy: History, measurement, and
theoretical perspectives. Handbook of personality: Theory and research, 2(1999), 102-
138.
Koren, Y., Bell, R., & Volinsky, C. (2009). Matrix factorization techniques for recommender
systems. Computer, 42(8), 30-37.
Koren, Y., & Bell, R. (2011). Advances in collaborative filtering. In Recommender Systems
Handbook (pp. 145-186). Springer US.
Kosinski, M., Stillwell, D., & Graepel, T. (2013). Private traits and attributes are predictable
from digital records of human behavior. Proceedings of the National Academy of
Sciences, 110(15), 5802-5805.
Lane, W., & Manner, C. (2011). The impact of personality traits on smartphone ownership and
use. International Journal of Business and Social Science, 2(17), 22-28.
Martin, D., Alzua, A., & Lamsfus, C. (2011). A contextual geofencing mobile tourism service.
In Information and communication technologies in tourism 2011 (pp. 191-202).
Springer Vienna.
Meehan, K., Lunney, T., Curran, K., & McCaughey, A. (2013). Context-aware intelligent
recommendation system for tourism. In Pervasive Computing and Communications
Workshops (PERCOM Workshops), 2013 IEEE International Conference on (pp. 328-
331). IEEE.
My Personality Project, (2014). start [myPersonality Project]. [online] Available at:
http://mypersonality.org/wiki/doku.php?id=start [Accessed 10 Sep. 2014].
Park, S. T., & Chu, W. (2009). Pairwise preference regression for cold-start recommendation.
In Proceedings of the third ACM conference on Recommender systems (pp. 21-28).
ACM.
Pazzani, M. J. (1999). A framework for collaborative, content-based and demographic filtering.
Artificial Intelligence Review, 13(5-6), 393-408.
Rentfrow, P. J., & Gosling, S. D. (2003). The do re mi's of everyday life: the structure and
personality correlates of music preferences. Journal of personality and social
psychology, 84(6), 1236.
Rentfrow, P. J., & Gosling, S. D. (2006). Message in a ballad the role of music preferences in
interpersonal perception. Psychological science, 17(3), 236-242.
Ricci, F., Rokach, L., & Shapira, B. (2011). Introduction to recommender systems handbook
(pp. 1-35). Springer US.
Rubens, N., Kaplan, D., & Sugiyama, M. (2011). Active learning in recommender systems. In
Recommender Systems Handbook (pp. 735-767). Springer US.
Shani, G., & Gunawardana, A. (2011). Evaluating recommendation systems. In Recommender
systems handbook (pp. 257-297). Springer US.
Shani, G., Chickering, M., & Meek, C. (2008). Mining recommendations from the web. In
Proceedings of the 2008 ACM conference on Recommender systems (pp. 35-42). ACM.
Tkalčič, M., Košir, A., & Tasič, J. (2013). The LDOS-PerAff-1 corpus of facial-expression
video clips with affective, personality and user-interaction metadata. Journal on
Multimodal User Interfaces, 7(1-2), 143-155.
Wang, Y., Chan, S. C. F., & Ngai, G. (2012). Applicability of Demographic Recommender
System to Tourist Attractions: A Case Study on Trip Advisor. In Proceedings of the
The 2012 IEEE/WIC/ACM International Joint Conferences on Web Intelligence and
Intelligent Agent Technology-Volume 03 (pp. 97-101). IEEE Computer Society.