Music Recommendations for Groups of Users Pedro Dias Dep. Computer Science Universidade Nova Lisboa Portugal p.dias@campus.fct.unl.pt João Magalhães Dep. Computer Science Universidade Nova Lisboa Portugal jm.magalhaes@fct.unl.pt ABSTRACT This paper presents an algorithm capable of providing mean- ingful recommendations to small sets of users. We consider not only rating patterns, bias tendencies, and temporal fluc- tuations, but also group-leaders. The approach here pre- sented intends to bring a fresh new look over group recom- mendations, making use of latent factor space to identify groups and make recommendations. Although these recom- mendations are oriented towards a few users, the preferences of their respective group leaders (users that better repre- sent the group) are also taken into account to diversify and smooth these recommendations. In contrast to the major- ity of group recommender systems described in literature, our system employs a collaborative filtering approach based on latent factor space instead of content-based or ratings merging approaches. Categories and Subject Descriptors H.3.3 [Information Search and Retrieval] Keywords Collaborative-filtering, groups, recommendations. 1. INTRODUCTION Recommender systems emerged with the purpose of pro- viding personalized and meaningful content recommenda- tions based on user preferences and usage history. Relying on the closest friends, family members or anyone else with whom one shares similarities to give trustworthy and useful advices has always been a characteristic of human behaviour, and different opinions weigh differently when it comes to making the final choice. The limitation on receiving good opinions from other people starts with the fact that, usually, one does not have many trustworthy or like-minded people to rely on for getting advice, and those few people have very limited knowledge, considering everything that exists and can be recommended, on a global scale point of view. Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full cita- tion on the first page. Copyrights for components of this work owned by others than ACM must be honored. Abstracting with credit is permitted. To copy otherwise, or re- publish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. Request permissions from permissions@acm.org. ImmersiveMe’13, October 22, 2013, Barcelona, Spain. Copyright 2013 ACM 978-1-4503-2402-1/13/10$15.00. http://dx.doi.org/10.1145/2512142.2512151. To explore the large number of ratings available on many online applications, we followed a collaborative filtering ap- proach to inter-relate and mine the relations between users and their preferences. Within collaborative filtering techniques, latent factor ap- proaches are very popular. The purpose of latent factor approaches to recommender systems is to map both users and products onto the same latent factor space, represent- ing these as vectors with k dimensions:  pu =(u1,u2, ··· ,u k ), qi =(i1,i2, ··· ,i k ) (1) Here, pu is the user u factors vector, qi is the product i factors vector and k is the number of latent factors (dimen- sions) upon which each user u and each product i are repre- sented. By representing users and products in such way, one can evaluate the extent to which users and products share common characteristics by comparing their k factors against each other. The principle underlying this approach is that both users and products can be represented under a com- mon reduced dimensionality space of latent factors that are inferred from the data and explain the rating patterns. Our algorithm operates exclusively in the latent-factor space. In the context of group recommendation, where there is more than one user to please, recommendations must be provided in a different way so that the whole group of users is satisfied. By operating in the latent-factor space one can easily relate different users. Moreover, by clustering this space we obtain a set of interest-groups to which users belong to. The leaders of these interest-groups are later used to broaden recommendations and cover products that satisfy all users. This paper is organized as follows: section 3 describes the matrix factorization implementation, section 4 presents the detection of leaders and the group-recommendation, and evaluation is detailed in section 5. Next, we discuss related work. 2. RELATED WORK Although recommender systems have recently attracted a lot of attention from the scientific community, group recom- mendation has not been widely addressed, since most recom- mendation techniques are oriented to individual users and focus on maximizing the accuracy of their preference pre- dictions. A. Jameson et al. [2] conducted an enlightening survey in 2007 presenting the most relevant works on the field of group recommendation, as well as the most com- mon issues addressed by the authors of the surveyed group recommender systems. The main challenges faced when pro- 21