Representação vetorial distribuída em sistemas de recomendação

• Ano de defesa: 2021
• Autor(a) principal: Pires, Pedro Reis
• Orientador(a): Almeida, Tiago Agostinho de
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de São Carlos; Câmpus Sorocaba
• Programa de Pós-Graduação: Programa de Pós-Graduação em Ciência da Computação - PPGCC-So
• Palavras-chave em Português: Sistemas de recomendação; Recommender systems; Filtragem colaborativa; Collaborative filtering; Modelos de embeddings; Embedding-based models
• Área do conhecimento CNPq: CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO
• Link de acesso: https://repositorio.ufscar.br/handle/ufscar/14225

Today, with the constant growth in the number of available digital content and ease of access to a huge amount of data, it is increasingly difficult to find relevant information. From this problem, recommender systems have emerged, analyzing user's behavior to issue personalized item recommendations. Currently, most companies have some form of content recommendation on their channels or services. Even so, many problems in the area are not properly solved, such as the high dimensionality and sparsity that the commonly adopted representation model has. Methods have been proposed to solve these problems, representing items and users as dense vectors in a space of reduced dimensionality. One of the most recent techniques in the literature is the use of embeddings-based models, i. e., distributed vector representations generated through artificial neural networks. Many of the latest advances in the area have shown promising results when compared to other already consolidated methods, but the vast majority of proposals suggest the execution of complex neural architectures or the use of content data, which often may not be available int the scenario of the application. In this dissertation, we extend the studies in the field of embeddings-based recommender systems. Through a comprehensive literature review, a new model for generating distributed representation is proposed, with the key points of being computationally efficient and requiring only implicit feedback from users to be trained. The results obtained in extrinsic and intrinsic evaluations, indicate that the model is promising for scenarios where there is need for computationally efficient models, achieving competitive results with state-of-the-art methods that demand greater computational power.