Identificação de uma vizinhança relevante em sistemas de recomendação para TV interativa utilizando tela secundária

• Ano de defesa: 2021
• Autor(a) principal: Ricardo Erikson Veras de Sena Rosa
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Minas Gerais
• Palavras-chave em Português: Engenharia elétrica; Sistemas de recomendação; Filtragem colaborativa; Multimídia interativa
• Link de acesso: https://hdl.handle.net/1843/37337

Advances in communication technologies have enabled the expansion of TV services, which has made it possible to make available an increasing amount of audiovisual content. This large amount of content introduces a problem known as information overload, where users cannot find the content of their interest in an efficient and timely manner. Recommender systems (RS) emerge as a promising tool to help users overcome information overload. This work explores two challenges for RS for Interactive TV: obtaining interaction data and improving the prediction accuracy for recommendations. The first challenge arises from the collective nature of TV environments, where interaction often occurs by using a remote control, which is shared by a group of users. As a result, it is difficult to identify the author of each interaction in a way that would be possible to customize the content for a specific user based on their individual interactions. The second challenge is inherent to RS. A greater accuracy translates into fewer recommendation errors, which increases users' confidence in using the system. To tackle these challenges, this work proposes, first, to use second screen devices to facilitate the capture of users' individual data. Since they are intended for personal use, interactions carried out using second screen devices can be individualized. To improve the prediction accuracy, this work seeks to find a more suitable neighborhood that relies on the similarity between users who evaluated groups of related objects, called local similarity. For this, the proposed method uses techniques that are based on clustering, resource allocation, and normalization. The evaluation of the proposed methods is performed by identifying and implementing scenarios for the individualized capture of interaction data and using prediction accuracy metrics on three databases widely used in the literature: MovieLens 100k, MovieLens 1M, and Netflix. The achieved results suggest the technical feasibility of the method for data capture using secondary screens and an improvement in the accuracy of the prediction approach that was proposed in this work.