Introduction
Human personality, as defined in psychology, is a combination of characteristics or
qualities that form an individual’s style of thinking, feeling and behaving in different
situations [
24
]. Among the existing models proposed to characterize and represent
human personality, the Five Factor model [
4
] states that there are five main factors
that allow describing an individual’s personality: openness, conscientiousness,
extraversion, agreeableness, and neuroticism.
Personality influences how people make their decisions [
18
]. Recent research has
shown that correlations between user personality traits and preferences exist in different
domains, such as music [
2, 20, 21, 22
], movies and TV shows [
2, 3, 19, 21
], books and
magazines [
2, 21
], and websites [
15
]. These correlations can be used to enhance
personalized information access and retrieval [
10, 17
]. Specifically, in recommender
systems, the exploitation of user personality information has enabled address cold-start
situations [
25
], facilitate the user preference elicitation process [
8
], mitigate the sparsity
problem [
11
], and improve the accuracy of collaborative filtering [
5, 23
].
Despite these achievements, the exploitation of user personality in recommender
systems is a challenging and still largely under explored topic.
Similarly to user preferences, personality factors can be inferred explicitly, e.g. by
means of psychometric questionnaires [
1, 4, 8
], or implicitly, e.g. by analyzing digital
footprints [
14
], linguistic features of user texts [
23
], and by correlating personality
traits with patterns of social network use, such as posting, rating, establishing
friendship relations, and participating in user groups [
1, 12
]. Whereas explicit
questionnaires are more accurate than techniques inferring personality from user
generated contents, they require the users’ effort to fill them. Note that the well
known IPIP [
7
] proxy for Costa and McCrae’s NEO-PI-R test [
4
] may have between
60 and 240 items. Hence, innovative techniques to efficiently acquire user
personality in recommenders have to be developed, and may be incorporated into the
user preference elicitation process [
8
].
Once extracted, it is needed to model user personality. Most of existing
personality-aware recommenders deals with a vector representation composed of the
numeric values of personality factors [
5, 8, 11
]. This, nonetheless, may not be the best
choice. There are works that have considered using discrete value intervals of the
personality factors [
5
], sets of predefined personality categories – e.g. reflective and
energetic people [
22
], and aesthetic, cerebral, communal, dark, and thrilling contents
[
21
] –, and personality-based stereotypes [
2, 16
]. In this context, a balance between
recommendation accuracy and comprehension (explicability) could be important.
Moreover, instead of broad user personality representations, using more fine-grained
information provided by the IPIP tests, such as the facets of each personality factor
[
9
] may help achieve better recommendations; Note e.g. that a particular individual
with a high overall openness score, may have high imagination and artistic interests,
but may not have a high level of adventurousness. Finally, the modeling task is even
more challenging if we account for variables that could influence particularities of an
individual’s personality. In this case, special attention should be given to the users’
age, gender, and educational attainment, as pointed out in [
2, 3, 5
].
Assuming that user personality can be inferred and modeled by a recommender
system, another set of open issues is related with the particular exploitation of user
personality for making recommendations. In general, simple approaches have been
investigated so far, from content-based [
23
] to collaborative filtering [
5, 8, 11, 17
]
heuristics. There is plenty of room for alternative, more sophisticated methods. In
particular, we envision matrix factorization as a powerful model in which user
personality information can be easily integrated with content-based and collaborative
filtering user profiles. Moreover, in addition to user preferences, contextual signals
may have an important role in personality-aware recommendations. In a particular
context, people with distinct personalities react differently. The identification and
exploitation of relations between context and personality is thus of special interest in
recommender systems. In fact, the user’s current mood is one of the contextual signals
that has been shown as directly related with personality [
18, 19
]. It is known that an
individual’s personality influences her mood changes due to external emotional
stimuli [
13
], and these stimuli may be generated by received recommendations. In all
the above cases, we may go beyond the accuracy of personality-aware
recommendations, and deal with other metrics, such as novelty and diversity [
20
];
Note, for example, that open-minded people may appreciate diverse and serendipitous
recommendations, while introverted people may prefer recommendations much more
related with their past preferences. Finally, we would like to highlight the fact that
exploiting personality could also help address new recommendation scenarios, such as
those of cross-domain recommender systems [
2
], in which information from a source
domain is used to enhance or generate recommendations in a different target domain,
where the user’s preferences may not be available [
6, 26
].