Towards Emotion-aware Recommender Systems: an Affective Coherence Model based on Emotion-driven Behaviors

Highlights

• An emotion-aware computational model based on affective user profiles is proposed

• An affective coherence score between an item and the user profile is defined

• The model is integrated in state-of-art recommendation approaches

• The way preferences depend on user emotional state varies from user to user

Abstract

Decision making is the cognitive process of identifying and choosing alternatives based on preferences, beliefs, and degree of importance given by the decision maker to objects or actions. For instance, choosing which movie to watch is a simple, small-sized decision-making process. Recommender systems help people to make this kind of choices, usually by computing a short list of suggestions that reduces the space of possible options. These systems are strongly based on the knowledge of user preferences but, in order to fully support people, they should be grounded on a holistic view of the user behavior, that includes also how emotions, mood, and personality traits influence her choosing patterns.

In this work, we investigate how to include emotional aspects in the recommendation process. We suggest that the affective state of the user, defined by a set of emotions (e.g., joy, surprise), constitutes part of choosing situation that should be taken into account when modeling user preferences.

The main contribution of the paper is a general emotion-aware computational model based on affective user profiles in which each preference, such as a 5-star rating on a movie, is associated with the affective state felt by the user at the time when that preference was collected.

The model estimates whether an unseen item is suitable for the current affective state of the user, by computing an affective coherence score that takes into account both the affective user profile and not-affective item features. The approach has been implemented into an Emotion-aware Music Recommender System, whose effectiveness has been assessed by performing in-vitro experiments on two benchmark datasets. The main outcome is that our system showed improved accuracy of recommendations compared to baselines which include no affective information in the recommendation model.

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:

• To build an affective user profile in which preferences are modeled considering affective information;

• 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.