Abordagens de Teoria dos Jogos para modelagem de Sistemas de Recomendação para grupos

• Autor(a) principal: Carvalho, Lucas Augusto Montalvão Costa
• Data de Publicação: 2013
• Tipo de documento: Dissertação
• Idioma: por
• Título da fonte: Repositório Institucional da UFS
• Texto Completo: https://ri.ufs.br/handle/riufs/3334
• topic: Inteligência artificial; Teoria dos jogos; Sistema de recomendação; Equilíbrio de Nash, Social Welfare; Artificial intelligence; Game theory; CNPQ::CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO

Recommendation systems have traditionally recommended items to individual users. In some scenarios, however, a recommendation for a group of individuals is necessary. The difficulty in performing recommendation for a group is how to properly deal with the preferences of its members to generate the recommendation. Different methods of aggregating these preferences have been proposed in the scientific literature, where the main goals are to maximize the average satisfaction of the group and ensure justice in the group recommendation. However, characteristics of the group greatly influence the results obtained by various aggregation methods. This paper defends the hypothesis that the Recommendation for Group of users can be modeled as a problem of finding the items in Nash Equilibrium. The items available for potential recommendation are modeled as actions of a Non-Cooperative Game. This approach selects items in a rational manner and treats members of the group as self-interested players. This ensures the existence of at least one Nash equilibrium as a solution to the group recommendation. The experiment compares the group average satisfaction between the proposed approach and some State of the Art aggregations strategies among them one known as Average. For groups of different levels of homogeneity, the results are very promising. Another hypothesis defended in this dissertation is that the formation of a group of users within a given context should be based on Alliance Structures with the goal of maximizing total Social Welfare of the group. While most recommender systems for groups recommend to a fixed group and predetermined user, groups organization can be performed according to a goal, for example, the suggestion of more homogeneous subgroups for better items recommendation for each of these subgroups. An experiment compared the outcome of the groups formation approach based on Alliance Structures with an approach based on a clustering method using K-Means algorithm. The results showed that the groups formed according to this new approach have an internal similarity index greater.