MovieOcean: Assessment of a Personality-based Recommender System
Luca Rolshoven
1
, Corina Masanti
1
, Jhonny Pincay
2,3 a
, Luis Ter
´
an
2
, Jos
´
e Mancera
2
and Edy Portmann
2 b
1
University of Bern, Hochschulstrasse 6, Bern, Switzerland
2
Human-IST Institute, University of Fribourg, Boulevard de P
´
erolles 90, Fribourg, Switzerland
3
Pontificia Universidad Cat
´
olica del Ecuador, Av. 12 de Octubre 1076, Quito, Ecuador
Keywords:
Recommender Systems, Personality-based Recommenders, Personalized Recommendations, Big Five Model.
Abstract:
This research effort explores the incorporation of personality treats into user-user collaborative filtering
algorithms. To explore the performance of such a method, MovieOcean, a movie recommender system
that uses a questionnaire based on the Big Five model to generate personality profiles, was implemented.
These personality profiles are used to precompute personality-based neighborhoods, which are then used
to predict movie ratings and generate recommendations. In an offline analysis, the root mean square error
metric is computed to analyze the accuracy of the predicted ratings and the F1-score to assess the relevance
of the recommendations for the personality-based and a standard-rating-based approach. The obtained
results showed that the root mean square error of the personality-based recommender system improves when
the personality has a higher weight than the information about the user ratings. A subsequent t-test was
conducted for the proposed personality-based approach underperformed based on the root mean square error
metric. Furthermore, interviews with users suggested that including aspects of personality when computing
recommendations is well-perceived and can indeed help improve current recommendation methods.
1 INTRODUCTION
While collaborative filtering methods are widely
used in implementing recommender systems (RS)
with good results, there is still room for improvement
in the quality of recommendations. The advent of
streaming platforms in the last few years (e.g.,
Netflix, Disney+, and Paramount+) and the easiness
of accessing a broad amount of media content have
led many companies and researchers to find methods
that provide better recommendations to engage users.
Because personality is about thinking, feeling, and
behaving, it seems coherent to think that it is possible
to improve the utility of recommendations by
harnessing personality data.
Authors such as Aaker (1999) had found that
brands that were associated with a set of personality
traits were perceived as more favorable when the
individuals were schematic on the personality
dimension that was also highly descriptive for the
a
https://orcid.org/0000-0003-2045-8820
b
https://orcid.org/0000-0001-6448-1139
brand. Mulyanegara et al. (2009) employed the Big
Five personality scale to assess the personality of 251
subjects in five dimensions and used those
assessments to measure the relationship between
consumer personality and brand personality in the
context of fashion products. Although the results
may differ in each domain, they suggest that
including personality aspects in the implementation
of RS engines could enhance the quality of
recommendations. The results of the research work
of Weaver Weaver III (1991) support the conception
that there is a correlation between personality traits
and media preferences across a variety of media.
This work addresses the following questions: (i)
how reliable are personality-based neighborhoods for
the computation of recommendations, and (ii) to
what extent are personality-based recommendations
perceived as more valuable than those based on
collaborative filtering? To answer these questions,
the authors implemented MovieOcean, a
personality-based movie recommender system that
uses the Big Five model to generate personality
profiles for each user. These profiles are used to
690
Rolshoven, L., Masanti, C., Pincay, J., Terán, L., Mancera, J. and Portmann, E.
MovieOcean: Assessment of a Personality-based Recommender System.
DOI: 10.5220/0011002500003179
In Proceedings of the 24th International Conference on Enterprise Information Systems (ICEIS 2022) - Volume 1, pages 690-698
ISBN: 978-989-758-569-2; ISSN: 2184-4992
Copyright
c
 2022 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved

precompute personality-based neighborhoods, then
used to predict movie ratings and generate
recommendations. Furthermore, the resulting
recommendations are assessed by conducting
experiments and measuring the root mean square
error (RMSE) and F1-score.
This article is structured as follows: Section 2
presents the theoretical background in which this
research effort is grounded. The methods developed
and applied in the execution of this work are
described in Section 3. Results are presented in
Section 4. Section 5 finishes the article with a
summary of the work conducted and the lessons
learned.
2 THEORETICAL BACKGROUND
Through the presentation of related work, this section
introduces the concepts and theories used to address
the above mentioned research questions.
2.1 Personality and the Big Five Model
The American Psychological Association (APA)
defines personality as follows (American
Psychological Association, 2020): “Personality
refers to individual differences in characteristic
patterns of thinking, feeling, and behaving. The study
of personality focuses on two broad areas:
Understanding individual differences in particular
personality characteristics, such as sociability or
irritability. The other is understanding how the
various parts of a person come together as a whole.”
When discussing personality and personality
characteristics in differential psychology, it is
essential to provide a shared taxonomy. The Big Five
personality model was developed as a result of the
findings of many study initiatives. It is made up of
five components (John et al., 2008):
• Openness: People with high openness tend to be
more original, open-minded, experimental, and
creative.
• Conscientiousness: People with high
conscientiousness tend to be more orderly,
responsible, goal-oriented, and organized.
• Extraversion: People with high extroversion
tend to be more energetic, enthusiastic, active,
sociable, and assertive.
• Agreeableness: People with high agreeableness
tend to be more altruistic, affectionate, modest,
trustful, and cooperative.
• Neuroticism: People with high neuroticism tend
to be more nervous, anxious, sad, and tense.
Which can be abbreviated as OCEAN.
The Big in Big Five was chosen to emphasize the
fact that these five factors are extensive and that they
were selected to represent personality at the broadest
level of abstraction (Pervin and John, 1999).
2.2 Personality Correlations
People differ significantly from each other based on
their personality traits. In psychology, researchers
aim to find the most defining characteristics detailed
enough to capture a personality but general enough
to avoid high complexity. Various studies have
proven a strong correlation and have also shown how
it can produce relevant predictions.
For instance, Rentfrow and Gosling (2003)
demonstrated how personality traits influence the
preferred music type. Also, Barrick and Mount
(1991) found out that there is a close relationship
between personality characteristics and job
performance. Other studies showed how personality
could influence the everyday online behavior of a
person. Preferences for specific social networks or
the amount of time that one spends online are also
correlated with the personality of the user, as Zhong
et al. (2011) were able to present in their manuscript.
Besides, Hu and Pu (2013) managed to show the
correlation between personality characteristics and
user rating behavior for retail products.
Cantador et al. (2013) looked at the relationship
between personality types and tastes in different
entertainment fields. Movies, TV shows, music, and
books were among the domains they selected. They
looked at data from 53 226 Facebook users to see any
significant links between personality traits and
preferences in the domains listed. In each of the
chosen domains, they generated personality-based
user stereotypes and association rules for 16 genres.
2.3 Other Related Work
In the literature, it is possible to find several
approaches to personality-based recommender
systems. Paiva et al. (2017) proposed a hybrid
recommendation approach that combines the Big
Five personality traits of the users with a correlation
between car fronts and power and sociability
perceptions to be used for semantic searches in
vehicle sales portals. They used the BFI-10 inventory
(Rammstedt and John, 2007), which solely consists
of 10 questions, to assess the Big Five traits of their
users. Then, they used the Euclidean distance
MovieOcean: Assessment of a Personality-based Recommender System
691

between users to form a neighborhood of size k = 3,
which is used with the correlations described above
to offer recommendations.
Braunhofer et al. (2015) used the Five-Item
Personality Inventory (FIPI), which includes even
fewer questions, to assess the personalities of the
users and then incorporate them into their
recommender system for Point of Interest (POIs)
recommendations. Their recommendation algorithm
makes use of matrix factorization (Koren et al.,
2009) and incorporates the demographic data of the
users as well as their personalities.
Another approach is to use a standard user-user
collaborative filtering algorithm and change the
similarity metrics to include personality information.
Tkal
ˇ
ci
ˇ
c et al. (2009) used this method and
determined that their personality-based similarities
led to an F1-score that was statistically equivalent to
the one obtained by using rating-based similarities.
However, they state that the personality-based
approach has three advantages: (i) an initial
questionnaire solves the new user problem, (ii) the
similarity computation between users is less
expensive, and (iii) the impact of the sparsity
problem is lowered. To assess the personalities of the
users, the authors used a version of the International
Personality Item Pool (IPIP) with 50 questions
(Goldberg et al., 2006).
In the realm of movies, Nalmpantis and Tjortjis
(2017) investigated the impact of integrating
personality into the recommendation process. They
used a hybrid approach to compute the
recommendations, in which a questionnaire was used
to measure the user’s personality characteristics. The
personality scores are then used together with the
findings of Cantador et al. (2013) to identify genres
that the consumer might enjoy. They estimated
movie ratings using collaborative filtering, for which
users had to rate 20 movies in advance. Finally, the
recommendations were generated by considering the
movie ratings and the user’s personality with a
weight of 0.5 each. They compared their system to
one that only considers recommendations based on
ratings and discovered that users preferred the
50%-50% approach to the rating-based method.
Considering previous research findings, using
personality characteristics to perform movie
suggestions seems to be a suitable way of enhancing
the quality and usefulness of the recommendations.
Compared to previous efforts, this work aims to
evaluate the effect of creating a personality-based
neighborhood and determine whether incorporating
personality traits increases the perceived value of the
recommendations among users.
3 METHOD AND USE CASE
The approach used in this study consists of three
major stages: (i) personality assessment, (ii)
recommendation engine design, and (iii) evaluation.
The following sections go into greater detail about
these phases and the implementation of MovieOcean.
Besides that, the movie information was obtained
from The Movie Database (TMBD)
1
through its free
developer API.
3.1 Personality Assessment
The personality of a user was determined with the
help of the Big Five model introduced in Section 2.1.
A standardized questionnaire containing 50
questions
2
from the (IPIP) was used. This
questionnaire is managed by the Oregon Research
Institute (Goldberg et al., 2006).
When the questionnaire is applied, each
participant can score up to 50 points in each
personality dimension (Openness,
Conscientiousness, Extraversion, Agreeableness, and
Neuroticism). The higher the score in a dimension,
the more the person can be described by this specific
personality trait. The particular questions consist of
statements that have to be rated on a 5-point Likert
scale from Very Inaccurate to Very Accurate. There
were precisely ten questions for each personality
dimension, some of them being positively keyed and
some of them being negatively keyed. For a question
that is positively keyed, the answer Very Inaccurate
would add one point to the corresponding personality
trait. In contrast, the answer Very Accurate would
result in five points added to that trait. For questions
that were negative keyed, the opposite order applies.
After completing the questionnaire, users could
see how many points they had obtained in each
personality dimension.
3.2 Recommendation Engine Design
To build the recommendation engine, a user-user
collaborative filtering algorithm was used.
Neighborhoods of size five were built for each user.
Neighborhoods were recalculated each time a new
user completed the IPIP questionnaire.
Recomputation of the neighborhoods was also
triggered if a user with an existing personality profile
decided to change one or more of their answers in the
questionnaire. Equation 1 was used to calculate the
similarity of two users a and k.
1
https://www.themoviedb.org/documentation/api
2
https://ipip.ori.org/new ipip-50-item-scale.htm
ICEIS 2022 - 24th International Conference on Enterprise Information Systems
692

sim(a, k) =
1
1 + p

sim
cos
(a, k) + p · sim
pers
(a, k)

(1)
The variable p is the personality factor used to
change the importance we want to give to the
personality in contrast to the ratings. In MovieOcean,
the value of p was set to 2, as the main focus was on
the personality more than on the ratings. Equation 2
was used for calculating the cosine similarity
between users a and k.
sim
cos
(a, k) =
∑
i∈I
r
k,i
· r
a,i
r
∑
i∈I
(r
k,i
)
2
·
r
∑
i∈I
(r
a,i
)
2
, (2)
where I is the set of movies that were rated by
both users and r
k,i
is the rating of user k for movie
i. The personality similarity sim
pers
(a, k) of the two
users can be calculated by using Equation 3:
sim
pers
(a, k) =
1
kφ(a) − φ(k)k + ε
(3)
In Equation 3, φ(a) corresponds to the personality
embedding, (i.e., a vector containing the scores for
each personality dimension of user a). Moreover, the
term ε is a small value used to avoid a division by
zero. The predicted rating of a movie i for user a is
then calculated using the standard prediction formula
in Equation 4.
r
a,i
= ¯r
a
+
∑
u∈N
a
(r
u,i
− ¯r
u
) · sim(a, u)
∑
u∈N
a
|sim(a, u)|
(4)
In the formula above, N
a
is a set of users in the
neighborhood of the active user a, ¯r
a
is the mean of
all the ratings issued by user a and ¯r
u
the mean rating
of user u respectively. This approach works as long
as some users already have an established personality
profile and ratings in the system. Otherwise, it
cannot deliver good results. This issue is known as
the cold start or new user problem. To address this
problem, movies with high ratings that were of a
genre well received by this particular personality
type were recommended to the user. To establish the
link between the personality and movie genres, the
results of Cantador et al. (2013) were used. Then the
three movie genres that had the least Euclidean
distance to the user’s personality were selected. The
user would then receive recommendations based on
these three genres and existing movie ratings, fetched
from TMDB.
3.3 Evaluation Method
The personality-based approach presented in this
work was compared with the standard rating-based
approach of user-user collaborative filtering to
evaluate the recommendation engine. Experiments
with different similarity functions and various values
for the neighborhood size and the personality factor
were conducted. Besides, it was verified how the
results changed by considering additional weighted
demographic data. To numerically capture which
approach performs better, two measurements were
applied: RMSE for assessing the accuracy of the
estimated ratings and the F1-score to measure the
relevance of the recommendations.
3.3.1 Similarity and Neighborhood
Experiments with different similarity functions and
different values for the neighborhood size were
conducted to find the best possible combination. For
the similarity function, four different methods were
chosen Choi and Suh (2013): the cosine similarity,
the Pearson correlation coefficient (PCC), the
inverted Euclidean distance, and the Jaccard index.
Our grid search procedure looked at neighborhoods
of sizes 5, 7, 9, 11, 13, and 15. In practice, the
neighborhood size is often larger than 15. In this
case, however, large neighborhoods were not used
because the dataset was small. A neighborhood of
fifteen users is already big compared to the total
number of users that we could include in the
evaluation. We first used the similarity function to
build the neighborhoods and then calculated the
predicted ratings of each user. In each step of the
grid search, one of the four similarity metrics was
used to build the neighborhood, either in the space of
user ratings or in the space of user personalities.
However, there was one exception: when looking at
the Jaccard Index, it was not used to build the
personality-based neighborhoods but only to
compute the rating similarity, which was still
considered in the personality-based predictions of the
ratings. Instead, the PCC was used to compute the
neighborhoods for the personality-based approach.
The Jaccard Index was not used to compute the
neighborhood of similar personalities because it does
not make sense in this scenario since the Jaccard
Index would always yield 1.
The following equations were used to compute
the predicted ratings, which are a modified version of
the standard collaborative filtering algorithm
(McLaughlin and Herlocker, 2004):
MovieOcean: Assessment of a Personality-based Recommender System
693

Rating: sim(a, k) =
1
1 + d

sim
rating
(a, k) + d · sim
dem
(a, k)

(5)
Personality:
sim(a, k) =
1
1 + p + d

sim
rating
(a, k) + p · sim
pers
(a, k) + d · sim
dem
(a, k)

(6)
In these equations, sim
rating
corresponds to the
rating similarity between two users, sim
pers
to the
similarity between the personalities of two users and
sim
dem
is the demographic similarity between the
users. The factors p and d correspond to the
personality factor and demographic factor. They are
used as a weight to give importance to the
personality or demographic data. In the first
experiment, p = d = 1 was set, and only the different
similarity metrics and neighborhood sizes were
changed. For sim
rating
and sim
pers
, the same
similarity that was applied to build the
neighborhoods was used.
The demographic similarity function sim
dem
was
the same for all experiments. It takes into account the
country of residence of the users, their age, and their
gender and is defined as follows:
sim
dem
(a, k) =
1
ˆc + ˆg + ˆa

ˆc · δ
c(a)c(k)
+ ˆg · δ
g(a)g(k)
+ ˆa ·
1
|a(a) − a(k)| + 1

(7)
In Equation 7, the following helper functions were
used: c(a): returns the country of user a; g(a): returns
the gender of user a; and a(a): returns the age of user
a.
Additionally, the Kronecker delta δ is used. It
returns 1 if the two values are equal (e.g.,
c(a) = c(k)) or 0 otherwise. The variables ˆc, ˆg, and ˆa
are the weights we give to each of the attributes. The
assumption behind this was that in some cases, it
might be important to give more importance to one
trait than another, especially if, for a given domain,
there exists knowledge about the strength of
correlations between certain demographic attributes
and the type of items that are used in the
recommender system. Finally, however, it was
decided to use the same weight for every
demographic attribute in our experiments (i.e.,
ˆc = ˆg = ˆa = 1).
3.3.2 Personality Factor
The second experiment used the best neighborhood
size and similarity measure that was found in the first
experiment. Note that these values might vary
depending on the approach that is being tested
(rating-based vs. personality-based). The goal of this
experiment was to determine the best personality
factor in the personality-based approach. The RMSE
and F1-score were computed for several personality
factors p, each time applying 3-fold cross-validation
for each user in the dataset. First, it was attempted to
look at personality factors with p = 0, 1, . . . , 10.
However, it was noticed that this magnitude of
change did not reflect itself much in the evaluated
metrics. Therefore, it was decided to use larger
personality factors with p = 0, 5, 10, . . . , 50. It is also
important to note that during this experiment, the
demographic factor d was set to 1.
3.3.3 Demographic Factor
The last experiment used different values for the
demographic factor d to see whether demographic
data could improve prediction accuracy and
recommendation relevance. The best configurations
found in the two previous experiments were applied,
and only the value for d was changed. More
precisely, the demographic factor was chosen to be
d = 0, 1, . . . , 10. The results of the three experiments
are presented in detail in section 4.
3.3.4 Interviews
Besides doing a quantitative analysis of the
personality-based collaborative filtering method,
opinions from the users of MovieOcean were
gathered. A small questionnaire was designed, and
interviews were conducted with three of the
registered users. Some of the questions asked
included: How did you find the questionnaire?, How
many questions do you find appropriate for such a
questionnaire in this setting?, Did you feel that it
captured your personality well?, and How accurate
were your recommendations?.
4 RESULTS
This section presents the results of implementing a
prototype following the methods explained in the
previous section.
ICEIS 2022 - 24th International Conference on Enterprise Information Systems
694

4.1 Prototype Implementation
A prototype of MovieOcean was implemented to
explore the incorporation of personality traits into a
recommender system. This artifact made it possible
to collect data from the users and test our
personality-based approach in movie
recommendations. In the following section, we will
give a brief overview of the implementation and
features of MovieOcean. The code of MovieOcean,
including its evaluation, is available in a public
repository on GitHub
3
.
In total, 42 users registered on the platform for
one month. To complete the registration process, the
users needed to answer a few questions about their
demographic background. These questions included
the age, the country where they lived, and the gender
of the user. Most of the users were aged between 20
to 30 years old and lived in Switzerland or Mexico.
The majority of the users were women, followed by
men, and three users picked another gender variant.
After the registration, the personality questionnaire
was presented to them. For the users that completed
the questionnaire, a personality profile that contains
the scores they reached in each personality
dimension was provided. Figure 1 presents an
example of the personality radar chart shown to the
users. Moreover, by issuing a few movie ratings, the
users had the chance to improve their
recommendations.
Figure 1: Screenshot of the personality scores that a user
would see after completing the questionnaire.
Each movie can be rated on a 5-star scale, and it
is possible to give half stars. The minimum rating
that can be issued is 0.5 stars. In total, 35 out of the
42 users completed the questionnaire and had,
therefore, a personality profile attached to their
account. However, only 28 users issued ratings. It
3
available at https://github.com/rolshoven/
MovieOcean/
was possible to collect a total of 445 movie ratings
during the studied period.
Figure 2: Comparison of the personality-based and rating-
based approach when using different similarity measures
and different neighborhood sizes.
4.2 Evaluation
The results of the first experiment are summarized in
the plots in Figure 2. It can be observed that the
metrics improved with a more significant number of
neighbors that were considered in the collaborative
filtering algorithm. The rating-based approach with
cosine similarity achieved the best RMSE of 1.694.
The best F1-score was achieved using the
personality-based strategy and the inverse Euclidean
distance as the similarity measure. The best
configuration was unambiguous for the rating-based
approach: cosine similarity and a neighborhood size
of 15 yielded the best results, both for the RMSE and
the F1-score. However, the results were more
ambiguous in the personality-based approach: The
best RMSE score was achieved using cosine
similarity and 15 neighbors. The best F1-score
occurred when using the inverse Euclidean distance
and 15 neighbors. We decided to use the inverse
MovieOcean: Assessment of a Personality-based Recommender System
695

Euclidean distance given that, in our opinion, the
F1-score is more relevant to the user than the
accuracy of the predicted ratings.
Regarding the second experiment, against our
initial assumptions, the personality factor did not
influence the F1-score of the personality-based
approach. However, it did improve the RMSE values
(as shown in Figure 3). It could be that an increasing
personality factor improves the accuracy for low or
high ratings but not much for ratings near the
threshold, which would result in better RMSE values
but the same F1-scores. On the left-hand side of
Figure 3, it can be observed that the RMSE
improvement of a more significant personality factor
decays exponentially. While increasing the
personality factor a lot does not influence the
computational complexity, we suggest not choosing a
too high value because it will limit the importance of
the rating similarity and demographic similarity. A
good value for p seems to be somewhere between 30
and 50. In our case, p = 50 was chosen before
starting the third experiment.
Figure 3: Comparison of the RMSE and F1-Scores using
different personality factors.
The last experiment was about the importance of
demographics. The results showed that considering
additional demographic data during the prediction of
movie ratings did not influence the performance of
the personality-based approach. We infer that a
correlation between personality scores and our
demographic attributes (i.e., country, gender, and
age) could explain these results. The personality
scores could then be seen as a proxy to the
underlying demographic data, therefore already
considering certain aspects of the demographics
implicitly. On the other hand, Figure 4 shows that the
demographics influenced the results of the
rating-based approach. When we did not consider
demographic data, the RMSE value for the
rating-based approach was 1.841, whereas it was
1.694 when we set d = 1, which is an improvement
of 0.147 or almost 8%. Our results suggest that
demographic data should be included in a
rating-based approach. However, the algorithm
should not give too much weight to demographic
similarities because it will not further improve the
performance. If the recommender system already
uses information about the user’s personality, there
seems to be no need to include additional
demographic information.
Figure 4: Comparison of the RMSE and F1-scores using
different demographic factors.
As a final comparison, a t-test on the RMSE and
F1-scores on the rating-based and personality-based
approach data in the third experiment with a
significance level of α = 0.05 was performed. It was
done in two scenarios: First, considering no
additional demographic data (d = 0) and second,
using additional demographic data (d = 1). The
results of this analysis are illustrated in Table 1.
There was no significant difference between the
F1-scores of the rating-based and the
personality-based method. Unfortunately, there was
a significant difference between the RMSE values of
the two approaches, with the personality-based
method performing poorer. The RMSE values for the
rating-based approach improved when using
additional demographic data, leading to an even
smaller p-value.
On the other hand, the personality-based
approach does have some advantages over the
rating-based approach. Suppose we know that the
F1-scores of these two approaches do not differ
significantly. In that case, we could assume that the
quality of recommendations will not worsen when
using the personality-based method.
Table 1: RMSE and F1-scores for the rating-based and
the personality-based approach with and without additional
demographic data. The last row shows the p-values of the
t-test that we performed.
Without demographics
(d = 0)
With demographics
(d = 1)
Approach RMSE F1-Score RMSE F1-Score
Rating-based 1.874 0.770 1.717 0.799
Personality-based 2.022 0.784 1.955 0.796
p-value 0.035 0.397 0.013 0.920
4.3 Interviews Outcome
The conception of recommending movies based on
the personality and the implementation of
MovieOcean was well-received by users. The design
ICEIS 2022 - 24th International Conference on Enterprise Information Systems
696

of MovieOcean was perceived as intuitive and
aesthetically pleasing. Only one out of the three
users had a problem finding a movie but also stated
that it could have been because of a typo in the
movie name. The users found it easy to navigate
through their recommendations and find other
movies that were linked as similar movies at the end
of each movie detail view. Moreover, the personality
assessment seemed to reflect what the user thought
about themselves. It was mentioned that the
overview of the personality scores and the short
explanations were received well. Another point is
that the star rating system was a good idea and
brought some life into the recommender system.
5 SUMMARY AND LESSONS
LEARNED
This work explored the inclusion of personality
characteristics as an enhancement of user-user
collaborative filtering algorithms. To this end,
MovieOcean, a personality-based movie
recommender system, was implemented.
MovieOcean uses a questionnaire based on the Big
Five model to generate personality profiles of the
users registered on the platform. With the personality
profiles, personality-based neighborhoods were
created to predict movie ratings and provide
recommendations. Furthermore, to assess how the
consideration of personality traits influences the
recommendations, the RMSE metric was calculated
to analyze the accuracy of the predicted ratings and
the F1-score to evaluate the relevance of the
recommendations for the personality-based and a
standard rating-based approach.
Unfortunately, it was not possible to find
evidence that incorporating personality in
collaborative filtering, in the way it was tested, leads
to better performance. The performance is
significantly worse in terms of the RMSE of the
predicted ratings. However, the F1-scores do not
differ significantly, which makes us think that the
recommendation quality is the same in both
approaches. If this is the case, it is still possible to
suggest using the personality-based approach
because it has several advantages as per our findings:
1. The neighborhoods of personalities can be
precomputed, which leads to a faster
recommendation process.
2. It is possible to use correlations between
personalities and movie genre tastes to overcome
the cold-start problem.
3. Users seem to enjoy the personality-based
approach, as the literature and interviews
conducted for this study showed.
Regarding the research questions defined in
section 1, we conclude that recommender systems
cannot rely more on a personality-based
neighborhood than on a traditional rating-based
neighborhood. However, given the case study results,
one could claim that it is valid to use
personality-based neighborhoods without
deteriorating the quality of the recommendations.
Moreover, only one method to incorporate
personality into recommender systems was explored.
The second question was about the accuracy of
the recommendations. Based on our findings, the
recommendation quality – measured by the F1-score
– does not differ significantly. The significantly
worse RMSE of the personality-based approach
could also influence the recommendations, but this
did not seem to be the case in the conducted
experiments. A slightly better F1-Score was found
for the personality-based approach, when no
additional demographic data was considered.
However, when taking demographic data into
account, the rating-based method had a slightly
higher F1-score. Table 1 shows that in both cases, the
difference was not significant. Suppose the F1-score
is more important than the RMSE of the predicted
ratings. In that case, the authors of this work believe
that it is possible to use a personality-based approach
without worrying about a negative effect on the
recommendation quality. However, based on the
obtained results, there are no improvements either.
In addition, the authors consider that the main
advantage of personality-based recommender
systems is that the neighborhood of similar users can
be precomputed, which results in a faster
recommendation process. In general, personalities
seem to be much more stable than rating vectors, and
we do not think the users would frequently change
the answers to their questionnaire. Even if they
change their answers, they probably only do so a few
times, resulting in a low number of recomputations
compared to the rating neighborhoods, which might
change very often. Moreover, the personality
information can be leveraged to solve the cold-start
recommendation problem by using correlations
between the user personalities and a prototype
personality of users that like movies of a specific
genre. While many users seem to like the
personality-based approach, there are probably also
other users that do not want to fill out the entire
questionnaire before using a recommender system.
In future work, the impact of different kinds of
MovieOcean: Assessment of a Personality-based Recommender System
697

Big Five questionnaires on the performance of a
personality-based recommender system will be
studied. Furthermore, other ways of including
personality aspects are going to be explored. For
instance, one approach would be to build the
neighborhoods using a standard rating similarity and
then only consider the users’ personality in the
similarity measure during the prediction of ratings.
REFERENCES
Aaker, J. L. (1999). The malleable self: The role of
self-expression in persuasion. Journal of marketing
research, 36(1):45–57.
American Psychological Association (2020). Personality.
https://www.apa.org/topics/personality. Accessed:
2020-09-19.
Barrick, M. R. and Mount, M. K. (1991). The big five
personality dimensions and job performance: a meta-
analysis. Personnel psychology, 44(1):1–26.
Braunhofer, M., Elahi, M., and Ricci, F. (2015). User
personality and the new user problem in a context-
aware point of interest recommender system. In
Information and communication technologies in
tourism 2015, pages 537–549. Springer.
Cantador, I., Fern
´
andez-Tob
´
ıas, I., and Bellog
´
ın, A. (2013).
Relating personality types with user preferences in
multiple entertainment domains. In CEUR workshop
proceedings. Shlomo Berkovsky.
Choi, K. and Suh, Y. (2013). A new similarity function
for selecting neighbors for each target item in
collaborative filtering. Knowledge-Based Systems,
37:146–153.
Goldberg, L. R., Johnson, J. A., Eber, H. W., Hogan,
R., Ashton, M. C., Cloninger, C. R., and 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.
Hu, R. and Pu, P. (2013). Exploring relations between
personality and user rating behaviors. In UMAP
Workshops.
John, O. P., Naumann, L. P., and Soto, C. J. (2008).
Paradigm shift to the integrative big five trait
taxonomy. Handbook of personality: Theory and
research, 3(2):114–158.
Koren, Y., Bell, R., and Volinsky, C. (2009). Matrix
factorization techniques for recommender systems.
Computer, 42(8):30–37.
McLaughlin, M. R. and Herlocker, J. L. (2004). A
collaborative filtering algorithm and evaluation metric
that accurately model the user experience. In
Proceedings of the 27th annual international ACM
SIGIR conference on Research and development in
information retrieval, pages 329–336.
Mulyanegara, R. C., Tsarenko, Y., and Anderson, A. (2009).
The big five and brand personality: Investigating
the impact of consumer personality on preferences
towards particular brand personality. Journal of brand
management, 16(4):234–247.
Nalmpantis, O. and Tjortjis, C. (2017). The 50/50
recommender: a method incorporating personality
into movie recommender systems. In International
Conference on Engineering Applications of Neural
Networks, pages 498–507. Springer.
Paiva, F. A., Costa, J. A., and Silva, C. R. (2017). A
personality-based recommender system for semantic
searches in vehicles sales portals. In International
Conference on Hybrid Artificial Intelligence Systems,
pages 600–612. Springer.
Pervin, L. A. and John, O. P. (1999). Handbook of
personality: Theory and research, chapter 4, pages
102–105. Guilford Press, 2 edition.
Rammstedt, B. and John, O. P. (2007). Measuring
personality in one minute or less: A 10-item
short version of the big five inventory in english
and german. Journal of research in Personality,
41(1):203–212.
Rentfrow, P. J. and 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.
Tkal
ˇ
ci
ˇ
c, M., Kunaver, M., Tasi
ˇ
c, J., and Ko
ˇ
sir, A.
(2009). Personality based user similarity measure for
a collaborative recommender system. In Proceedings
of the 5th Workshop on Emotion in Human-Computer
Interaction-Real world challenges, pages 30–37.
Weaver III, J. B. (1991). Exploring the links between
personality and media preferences. Personality and
individual differences, 12(12):1293–1299.
Zhong, B., Hardin, M., and Sun, T. (2011). Less
effortful thinking leads to more social networking?
the associations between the use of social network
sites and personality traits. Computers in Human
Behavior, 27(3):1265–1271.
ICEIS 2022 - 24th International Conference on Enterprise Information Systems
698