Towards Emotion-aware Recommender Systems: an Affective Coherence Model based on Emotion-driven Behaviors
Introduction
The great availability of personalized services over the Internet has made users more inclined to provide their data in order to obtain accurate suggestions about products to buy, news to read, music to listen to. Today, companies such as Netflix or Amazon collect millions of records about habits of their customers and use them to target their suggestions by exploiting recommender systems (RecSys), that are filtering tools that guide the user in a personalized way to interesting or useful items in a large space of possible options (de Gemmis et al., 2017; Burke, 2002, Narducci et al., 2020). These systems collect information about user preferences, either explicitly by asking users to provide ratings on items, or implicitly by analyzing their actions on items (e.g., downloads, prints, views). Collected data are exploited to build a model of user preferences (user profile), or to discover users having similar interests, with the aim of finding novel items that might be interesting to them. Today, personalization systems could take advantage of digital footprints, i.e. the entire collection of information generated by a person’s online activity, such as online searches, purchased items, or posts on social media. In particular, actions performed on Social Media Sites (SMSs) provide valuable information about user tendencies, styles of life, as well as affective and psychological traits (Ryan et al., 2011, Asur et al., 2010, Correa et al., 2010, Malhotra et al., 2012). It is not surprising that social media footprints, such as liked pages, attended events, shared music, etc., are being more and more exploited by RecSys (Kazai et al., 2016, Ma et al., 2017) to infer personality traits (Back et al., 2010), affective states or tendencies (Alm et al., 2005, Mohammad and Kiritchenko, 2015, Polignano et al., 2017), with the aim of improving accuracy of suggestions. Recently, the European Legislation has strongly limited the use of that data without an explicit agreement by the final user. In particular, the GDPR regulation (Goddard & Michelle, 2017) has placed restrictions on the exploitation of user data in order to limit illegal use. In our work, we used only third-party anonymized publicly available datasets. In the case the system will be implemented in real-world environments, the privacy issue should be addressed, of course.
Several studies in Psychology described how emotions actively interact with the choosing patterns evoked in the subject during decision processes (Picard et al., 2004, Clore et al., 1994, Bechara et al., 2000, Loewenstein and George, 2003, Frijda et al., 1994, Cunningham and Michael R, 1988) and showed that those areas of our brain related to feelings are highly stimulated during decision-making processes (Bechara & Antoine, 2003).
Therefore, novel techniques to model user preferences and attitudes, according to a holistic view of personalization, are needed. The main contribution of this work is the investigation of the following open issues:
- 1.
To build an affective user profile in which preferences are modeled considering affective information;
- 2.
To include an affective profile into a recommendation process.
As a proof-of-concepts, we implemented an Emotion-aware RecSys and performed an evaluation in the music domain. The results showed improved accuracy of recommendations compared to several baselines, including machine learning methods and content-based filtering methods which exploit no affective information to model user preferences. The rest of the paper is organized as follows: Section 2 provides an analysis of the main relevant work for this research, Section 3 describes our original model for generating emotion-aware recommendations, implemented in EMRES that is described in Section 4. Finally, the experimental evaluation and the results are outlined in Section 5.
Section snippets
Related Work
The proposed investigation falls at the intersection between Psychology and Computer Science. In this section, we provide some background and related work that allow to place our research in the appropriate context of both areas. In the analysis of Psychology literature, we focus on the role that affect (emotions, mood, personal feelings) could play in decision-making processes, in order to motivate the design choices behind our affect-aware model. Then, we describe some recent attempts to
An Emotion-aware Computational Model for Personalization Systems
We propose a general model that could be adopted by personalization systems that exploit affective information to better tailor services to users. In Fig. 1, a user-centered computational model is proposed, which takes into account user tendencies, behaviors, and preferences. The left side of the picture describes the part of the model devoted to perform data collection: digital footprints left on social networks (posts, tweets, reviews, etc.) and user behavior, observed when interacting with
EMRES: Emotion-aware Music REcommender System
EMRES computes a list of suggested songs for user u, by taking into account his affective profile. The part of the model that performs data collection is implemented by defining:
- 1.
a strategy for collecting user affective states and preferences on songs;
- 2.
a model to represent songs and to compute song similarity.
Fig. 2 shows the general model of EMRES and the main procedural steps for obtaining recommendations. In particular, starting from point 1.1 of the training phase, it is possible to notice
Experimental Evaluation
The main aim of the evaluation is to compare our emotion-aware RecSys EMRES with alternative approaches which do not take into account affective information, in order to show that the adoption of the proposed Affective Coherence Model improves the accuracy of suggestions. As alternative approaches we consider: (i) a content-based filtering algorithm that exploits the same non-affective features as our system; (ii) a recommendation method based on machine learning techniques for rating
Conclusion
Traditional recommendation approaches have started to include novel aspects that describe in a deeper way the context in which a decision is taken, in order to increase both accuracy and the acceptance of proposed suggestions. In this work, we investigated how to model the affective dimension of the user in a recommendation process, based on the idea that preferences of users might vary according to what they feel.
The main contribution of this work is a model that computes the affective
Acnowledgments
The research has been partially funded by the Apulia Region, Italy, POR Puglia FESR-FSE Innonetwork, project DECiSION: DATA-DRIVEN CUSTOMER SERVICE INNOVATION.
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
References (82)
- et al.
Who interacts on the Web? The intersection of users’ personality and social media use
Computers in Human Behavior
(2010) - et al.
An Investigation on the Serendipity Problem in Recommender Systems
Information Processing & Management
(2015) - et al.
Exploring user emotion in microblogs for music recommendation
Expert Systems with Applications
(2015) A naturalistic decision making perspective on studying intuitive decision making
Journal of Applied Research in Memory and Cognition
(2015)- et al.
Efficient L1 Regularized Logistic Regression
- et al.
Modeling and broadening temporal user interest in personalized news recommendation
Expert Systems with Applications
(2014) - et al.
Universal affective model for Readers’ emotion classification over short texts
Expert Systems with Applications
(2018) - et al.
Sentiment analysis algorithms and applications: A survey
Ain Shams Engineering Journal
(2014) - et al.
SCoR: a synthetic coordinate based recommender system
Expert Systems with Applications
(2017) - et al.
The function of concrete emotions in rational decision making
Acta Psychologica
(1992)
EARS: Emotion-aware recommender system based on hybrid information fusion
Information Fusion
Who uses Facebook? An Investigation into the Relationship between the Big Five, Shyness, Narcissism, Loneliness, and Facebook Usage
Computers in Human Behavior
The role of user mood in movie recommendations
Expert Systems with Applications
Patterns of temporal variation in online media
A Comparison of General vs Personalised Affective Models for the Prediction of Topical Relevance
Facebook profiles reflect actual personality, not self-idealization
Psychological Science
A survey on context-aware systems
IJAHUC
Risky business: emotion, decision-making, and addiction
Journal of Gambling Studies
Emotion, decision making and the orbitofrontal cortex
Cerebral Cortex
Context-aware points of interest suggestion with dynamic weather data management
Hybrid Recommender Systems: Survey and Experiments
User Modeling User-Adapted Interaction
Personality and Recommendation Diversity
Affective causes and consequences of social information processing
Handbook of Social Cognition
What do you do when you’re happy or blue? Mood, expectancies, and behavioral interest
Motivation and emotion
Universals and cultural differences in the judgments of facial expressions of emotion
Journal of Personality and Social Psychology
The EU General Data Protection Regulation (GDPR): European regulation that has a global impact
International Journal of Market Research
Music emotion classification and context-based music recommendation
Multimedia Tools and Applications
Cited by (48)
Emotion-oriented recommender system for personalized control of indoor environmental quality
2024, Building and EnvironmentERNetCL: A novel emotion recognition network in textual conversation based on curriculum learning strategy
2024, Knowledge-Based SystemsTargeted aspect-based emotion analysis to detect opportunities and precaution in financial Twitter messages
2023, Expert Systems with ApplicationsVirtual Customer Assistants in finance: From state of the art and practices to design guidelines
2023, Computer Science ReviewHELENA: An intelligent digital assistant based on a Lifelong Health User Model
2023, Information Processing and ManagementCitation Excerpt :User profiling is nowadays a common practice over the web, which began with seminal work by Kobsa, Koenemann, and Pohl (2001), who proposed a keyword-based representation of the web user. Since then, many strategies have been proposed to exploit the “digital footprints” that people leave on the web (i.e., social networks, web services, purchases, preferences, searches, etc.) to build a user profile (Polignano, Basile, Rossiello, de Gemmis, & Semeraro, 2017; Polignano, Gemmis, Narducci, & Semeraro, 2017; Polignano, Narducci, de Gemmis, & Semeraro, 2021; Song, Lee, & Moon, 2019), a representation of his/her interests and preferences to be used for personalization and adaptation purposes. Among possible user profile facets (Cena, Likavec, & Rapp, 2019), we focus on the physical and mental state.