Exploiting linked data in Dbpedia to reduce prediction error in matrix factorization recommenders

• Ano de defesa: 2022
• Autor(a) principal: Pereira, Victor Martinez Vidal
• Orientador(a): Durão, Frederico Araujo
• Banca de defesa: Durão, Frederico Araujo , Pereira, Adriano César Machado , Coimbra, Danilo Barbosa
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Universidade Federal da Bahia
• Programa de Pós-Graduação: Programa de Pós-Graduação em Ciência da Computação (PGCOMP)
• Departamento: Instituto de Computação - IC
• País: Brasil
• Palavras-chave em Português: Sistemas de recomendação; Fatorização de matrizes; Dados abertos; Erro predito
• Área do conhecimento CNPq: CNPQ::CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO::METODOLOGIA E TECNICAS DA COMPUTACAO::SISTEMAS DE INFORMACAO
• Link de acesso: https://repositorio.ufba.br/handle/ri/36099

Recommender Systems provide suggestions for items that are most likely of interest to users. Providing personalized recommendations is a challenge that can be addressed by filtering algorithms among which Collaborative Filtering (CF) has demonstrated much progress in the last few years. By using Matrix Factorization (MF) techniques, CF methods reduce prediction error by using optimization algorithms. However, they usually face problems such as data sparsity and prediction error. Studies point to the use of data available in Semantic Web as a path to improve recommender systems and address the challenges related to CF techniques. Motivated by these premises, the present work, conducted by me at RecSys Research Group at UFBA, developed a data pipeline along with an algorithm that processes the Ratings Matrix combining semantic similarities of Linked Open Data (LOD) and estimates missing ratings. The experiments took subsets of 1000 samples from three di↵erent datasets (Movielens, LastFM and LibraryThing), calculated two semantic similarity metrics, Linked Data Similarity Distance (LDSD) and Resource Similarity (RESIM), and applied three MF-based algorithms (SVD, SVD++ and NMF). Results suggest the proposed pipeline is able to reduce Root Mean Square Error (RMSE) of all subsets with statistical confidence supported by parametric test one-way ANOVA followed by Tukey’s multiple comparison test.