GUARD: um arcabouço para recomendação baseado em programação genética

• Ano de defesa: 2013
• Autor(a) principal: Adolfo Pinto Guimarães
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Minas Gerais; UFMG
• Palavras-chave em Português: Ciência da Computação; Computação; Sistemas de recuperação da informação; Programação genética (computação)
• Link de acesso: http://hdl.handle.net/1843/ESBF-97GR3N

Recommender systems suggest new items to the user based on his/her interest. These systems appear in distinct contexts, including e-commerce, search engines, program guides for digital TV. This work proposes GUARD (A Genetic Unified Approach for Recommendation),a framework based on genetic programming conceived to generate items ranking functions for recommendation. The framework is flexible, and although developed under a colaborative filtering framework can be easily extended to work with content-based or hybrid recommender systems. When working with collaborative filtering, items are recommended to the user based on preferences of similar users in the system. Genetic Programming is a method based on the theories of evolution and survival of the fittest. The main motivation behind using genetic programming to generate items raking functions is in their flexibility to combine different data evidences and capability of dealing with data uncertainty and noise. GUARD evaluates the generated ranking functions using four different measures: precision, recall, diversity and novelty. The evaluation, which can be based any combination of these criteria, follows two different approaches for multicriteria optimisation: a Pareto-based and a lexicographical approach. The framework was tested in the scenario of movies recommendation, using the Movielens 100k and 1M datasets. The results obtained were compared to those generated by PureSVD, the state-of-the-art algorithm for collaborative filtering. Considering the Movielens 100k, the results of precision and recall were superior to those of SVD. The framework can also generate more diverse and novel recommendations, with a small loss in precision. For Movielens 1M, the results are not better than those of PureSVD. The generated ranking functions lose in accuracy for PureSVD but gain in simplicity and efficiency..