Choice-based recommender systems Paula Saavedra CITIUS University of Santiago de Compostela Santiago de Compostela, Spain paula.saavedra@usc.es Pablo Barreiro CITIUS University of Santiago de Compostela Santiago de Compostela, Spain pablobv70@gmail.com Roi Durán CITIUS University of Santiago de Compostela Santiago de Compostela, Spain roiduram@gmail.com Rosa Crujeiras School of Mathematics University of Santiago de Compostela Santiago de Compostela, Spain rosa.crujeiras@usc.es María Loureiro School of Business University of Santiago de Compostela Santiago de Compostela, Spain maria.loureiro@usc.es Eduardo Sánchez Vila CITIUS University of Santiago de Compostela Santiago de Compostela, Spain eduardo.sanchez.vila@usc.es ABSTRACT Choice-based models are proposed to overcome some of the limitations found in traditional rating-based strategies. The new approach is grounded on decision-making paradigms, such as choice and utility theories. Specifically, random utility models were applied in a recommendation problem. Prediction accuracy was compared with state-of-art rating- based algorithms in a gastronomy dataset. The results show the superior performance of choice-based models, which may suggest that real choices could bring more predictive power than ratings. CCS Concepts •Information systems → Collaborative filtering; So- cial recommendation; Keywords Choice models; Random Utility Models; Logit probabilities; Tourism 1. INTRODUCTION Recommender systems are personalization tools aimed at suggesting relevant items on the basis of available informa- tion on items as well as decision-makers [5]. Broadly speak- ing, recommenders can be classified in two different cate- gories. Content-based recommenders generate a profile for each decision-maker by considering items experienced in the past. The profile typically represents the preferences of the decision-maker, i.e the taste of the decision-maker on each Copyright held by the author(s). RecTour 2016 - Workshop on Recommenders in Tourism held in conjunc- tion with the 10th ACM Conference on Recommender Systems (RecSys), September 15, 2016, Boston, MA, USA. item’s attributes [2]. These preferences can be used to pre- dict the utility of any given item by comparing them with the values of item’s attributes. Collaborative recommenders, on the other hand, take advantage of previous ratings provided by the available decision-makers to predict the utility of any given user-item pair [6]. This approach has been widely adopted as it removes the burden of knowing and managing item attributes as well as their corresponding values. Many algorithms and models have been proposed under the collaborative paradigm. Among them, two families have gained major attraction: neighborhood algorithms and la- tent factor models. The neighborhood approach was the first to implement to collaborative concept and became the reference model in this research area [9, 4]. The method con- sists on representing vectors of ratings on either the decision- maker or item space. The distance between any pair of these vectors determine the similarity between either the decision- makers or the items that these vectors represent. Individuals with similar rating’s vectors are considered to possess similar tastes or preferences, while items are considered to have sim- ilar attributes. The latent factor strategy, in turn, attempts to explain ratings by means of characterizing both users and items with a limited set of factors. Factors are considered unknown variables that can be inferred from the ratings de- clared by the users. The inference or learning problem can be solved with factorization techniques. The classical fac- torization method is called Singular Value Decomposition (SVD) and was applied successfully to identify and reduce the number of relevant factors [10]. However, the method requires complete knowledge of the rating matrix and fill-in methods to populate sparse rating matrix come at a cost of inaccurate factor learning. Recently, new factorization tech- niques have been successfully developed that are capable of learning the factors from sparse rating matrices [7]. Each rating is explained by means of two vectors whose dimen- sions correspond with the set of latent factors. The first vector represents the item in terms of its degree of posse- sion of each factor, while the second vector represent the decision-maker on the basis of her preference on each factor. These item and decision-maker vectors constitute a pair of