Most of the research in recommender systems focuses on data-driven approaches. In this paper we present our vision for complementing data-driven approaches with model-driven ones. We present a preliminary experimental set-up and we expose our research plan. In the experimental set-up we acquired eudaimonic characteristics of movies and user preferences. Furthermore, we performed a preliminary analysis of the acquired data.
Mainstream research in recommender systems is data-driven. Logs of user
behaviour, such as clicks, ratings, purchases, are used in a variety of algorithms,
such as collaborative and feature-based approaches to generate
recommendations [
We view recommender systems as tools for helping users make better
decisions [
We assume that users di er in their need for eudaimonic experiences, i.e. some people prefer to just have fun, while other people may prefer to spend their time contemplating meaning and purpose. Similarly, we observe that movies di er in the experience quality they induce. For example, the movies The Hangover and La vita e' bella are both comedies. But while the former is a shallow comedy with a series of simple jokes the latter deals with deeper issues, such as the holocaust.
In positive psychology, happiness is often described through two opposite
concepts: hedonism and eudaimonism [
In order to devise personalization approaches using eudaimonia there are a lot of steps to make, since it is an unexplored area. We foresee the following steps need to be taken: 1. unobtrusive inference of eudaimonic and hedonic user preferences 2. automatic labeling of movies' eudaimonic and hedonic qualities 3. a personalized recommender system that takes advantage of eudaimonic and hedonic features 4
We performed a user study to acquire data. We let the subjects choose movies
from a pool of popular movies. For the hypothetical context of choosing a movie
to watch alone on a Saturday evening, each subject had to choose the most
appropriate movie (the liked movie) and the least appropriate movie (the disliked
movie). The subjects were then asked to describe, for each of the two movies,
their viewing experience in terms of hedonic and eudaimonic experience using
an adaptation of the scale developed by [
We hand-picked the movies in order to have a mix of hedonic and eudaimonic movies.
We ran the study through Amazon Mechanical Turk. After removing subjects who did not pass a control question and removing outliers using the Mahalanobis distance we had the answers of 84 subjects (M = 34:2 years, SD = 9:5 years, 29 females). 5
A deeper analysis of the data acquired is going to be presented at the UMAP
2018 conference [
Users divided themselves into two clusters in terms of the eudaimonic qualities of the liked movies (see Fig. 1): some users liked movies with high eudaimonic qualities (scores > 3) while some users liked movies with low eudaimonic qualities (scores < 3), which is re ected in the bimodal shape of the histogram in Fig. 1. We conjecture that this bi-modal shape is due to user being either pleasure-seekers or meaning-seekers.
We clustered the movies into three categories: hedonic-only, eudaimonic-only and mixed. We performed the Wilcoxon rank sum test in order to test the hypothesis of the mean reported hedonic and eudaimonic quality being equal. Examples from all three clusters are reported in Tab. 2.
I liked the movie because it challenged my way of seeing the world. 30 20 t n u o c 10 0 1 2 dfClean$3eud_1_1 4 5 The results of our study indicate that eudaimonic characteristics are useful for accounting variance in user preferences and for characterizing movies.
We plan to proceed the with the three steps: (i) unobtrusive inference of eudaimonic/hedonic preferences, (ii) automatic labeling of movies and (iii) personalized recommendations of movies.
For the unobtrusive inference we are designing an experiment. In addition to the variables acquired in the experiment reported above, we will collect also the users' social media data. We plan to ask for Facebook likes, twitter activity and Instagram activity. Using features extracted from social media activity we plan to do a predictor of the user preferences for movies in the hedonic/eudaimonic space. The recent scandal with Facebook data integrity may pose a further probblem in devising such a method.
For the automatic labeling of movies we plan to use movie subtitles for feature generation. We foresee the usage of NLP techniques and generate features using TF-IDF and embeddings. In order to get ground truth movie labels we plan to crowd-source the labeling of a pool of movies.
For the personalized recommendation part we plan to use content-based recommendation methods that take advantage of the eudaimonic and hedonic features.