Feature selection for movie recommendation

Abstract

TV users have an abundance of different movies they could choose from, and with the quantity and quality of data available both on user behavior and content, better recommenders are possible. In this paper, we evaluate and combine different content-based and collaborative recommendation methods for a Turkish movie recommendation system. Our recommendation methods can make use of user behavior, different types of content features, and other users' behavior to predict movie ratings. We gather different types of data on movies, such as the description, actors, directors, year, and genre. We use natural language processing methods to convert the Turkish movie descriptions into keyword vectors. Then, for each user, we use the content features and the user's past implicit ratings to produce content feature-based user profiles. In order to have more reliable profiles, we do feature selection on these profiles. We show that for each feature space, such as actor, director, or keyword, a different amount of feature selection may be optimal. Different recommenders may also perform best for a different number of movies available as training data for a user. We also combine different content-based recommenders and collaborative recommenders using an aggregation or the best of the available recommenders. Experimental results on a dataset with hundreds of users and movies show that, especially for users who have watched a small number of movies in the past, feature selection can increase recommendation success.