Tell me what you want: Embedding narratives for movie recommendations

Recommender systems are efficient exploration tools providing their users with valuable suggestions about items, such as products or movies. However, in scenarios where users have more specific ideas about what they are looking for (e.g., they provide describing narratives, such as "Movies with minimal story, but incredible atmosphere, such as No Country for Old Men"), traditional recommender systems struggle to provide relevant suggestions. In this paper, we study this problem by investigating a large collection of such narratives from the movie domain. We start by empirically analyzing a dataset containing free-text narratives representing movie suggestion requests from reddit users as well as community suggestions to those requests. We find that community suggestions are frequently more diverse than requests, making a recommendation task a challenging one. In a prediction experiment, we use embedding algorithms to assess the importance of request features including movie descriptions, genres, and plot keywords, by computing recommendations. Our findings suggest that, in our dataset, positive movies and keywords have the strongest, whereas negative movie features have the weakest predictive power. We strongly believe that our new insights into narratives for recommender systems represent an important stepping stone towards novel applications, such as interactive recommender applications.