Finding Similar Movies: Dataset, Tools, and Methods

Abstract

Recommender systems are becoming ubiquitous in online commerce as well as in video-on-demand (VOD) and music streaming services. A popular form of giving recommendations is to base them on a currently selected product (or items), and provide "More Like This," "Items Similar to This," or "People Who Bought This also Bought" functionality. These recommendations are based on similarity computations, also known as item-item similarity computations. Such computations are typically implemented by heuristic algorithms, which may not match the perceived item-item similarity of users. In contrast, we study in this paper a data-driven approach to similarity for movies using labels crowdsourced from a previous work. Specifically, we develop four similarity methods and investigate how user-contributed labels can be used to improve similarity computations to better match user perceptions in movie recommendations. These four methods were tested against the best known method with a user experiment (n = 114) using the MovieLens 20M dataset. Our experiment showed that all our supervised methods beat the unsupervised benchmark and the differences were both statistically and practically significant. This paper's main contributions include user evaluation of similarity methods for movies, user-contributed labels indicating movie similarities, and code for the annotation tool which can be found at http://MovieSim.org.