Blending Users, Content, and Emotions for Movie Recommendations

Abstract

Recommender systems were initially deployed in eCommerce applications, but they are used nowadays in a broad range of domains and services. They alleviate online information overload by highlighting items of potential interest and helping users make informed choices. Many prior works in recommender systems focussed on the movie recommendation task, primarily due to the availability of several movie rating datasets. However, all these works considered two main input signals: ratings assigned by users and movie content information (genres, actors, directors, etc). We argue that in order to generate high-quality recommendations, recommender systems should possess a much richer user information. For example, consider a 3-star rating assigned to a 2-hour movie. It is evidently a mediocre rating meaning that the user liked some features of the movie and disliked others. However, a single rating does not allow to identify the liked and disliked features. In this talk we discuss the use of emotions as an additional source of rich user modelling data. We argue that user emotions elicited over the course of watching a movie mirror user responses to the movie content and the emotional triggers planted in there. This implicit user modelling can be seen as a virtual annotation of the movie timeline with the emotional user feedback. If captured and mined properly, this emotion-annotated movie timeline can be superior to the one-off ratings and feature preference scores gathered by traditional user modelling methods. We will discuss several open challenges referring to the use of emotion-based user modelling in movie recommendations. How to capture the user emotions in an unobtrusive manner? How to accurately interpret the captured emotions in context of the movie content? How to integrate the derived user modelling data into the recommendation process? Finally, how can this data be leveraged for other types of content, domains, or personalisation tasks?