LinearTag Models: Recommendations Using Linear User Profiles Based on Tags

Abstract

Recommender systems allow the exploration of large collections of products, the discovery of patterns in the products, and the guidance of users towards products that match their interests. Collaborative tagging systems allow users to label products in a collection using a free vocabulary. The aggregation of these tags, also called a Folksonomy, can be used to build a collective characterization of the products in a simple and recognizable vocabulary. In this paper, we propose a family of methods called LinearTag recommenders, which infer users preferences for tags to formulate recommendations for them. We dubbed these inferred user profiles as TagProfiles. We present experiments using them as an interaction artifact that allows users to receive new recommendations as they delete, add or reorder tags in their profiles. Additional experiments using the Movielens dataset, show that the proposed methods generate recommendations with an error margin similar, or even lower than the results reported by methods based on latent factors. Next, we compared TagProfiles against KeywordProfiles, which are profiles based on keywords extracted automatically from textual descriptions of products. This comparison showed that TagProfiles are not only more precise in their predictions, but they are also more understandable by users. At last, we developed a user interface of a movie recommender based on TagProfiles, which we tested with 25 users. This experience showed that TagProfiles are easier to understand and modify by users, allowing them to discover new movies as they interact with their profiles.