Recomendación de productos a partir de perfiles de usuario interpretables

Abstract

Recommender systems allow users to have a personalized view of large sets of products, relieving the overload problem of choice in e-commerce sites. Usually, recommendations are obtained using the technique called "collaborative filtering". This technique filters the products the users wish, from those they don´t want, inferring affinities between products and users in a space of abstract features, also called a latent space. These techniques have proven to be of great predictive value, but these created profiles are neither understandable, nor editable for users, enclosing users in a bubble, in which they only receive collaborative recommendations conditioned by their historical behaviors. In our work we propose a method to build user profiles, defined in interpretable spaces, or defined in terms of collaborative tags or keywords (i.e. words extracted from the descriptions of the product), which can be interpreted and modified by users. The model proposed generate linear profiles, whose coefficients, positives or negatives, reflect the user's affinity towards tags or keywords, according to the space selected. To test our hypothesis, we used the dataset of research in movie recommender systems from the University of Minnesota: Movielens. The results show that the predictive ability of the model, based on interpretable user profiles, is comparable to those models based on abstract profiles with the added benefit that these profiles are interpretable.

Keywords

• Collaborative filtering
• collaborative tagging systems
• recommender systems
• social tagging
• user interfaces