Content recommendation based on semantic descriptions extracted from user annotations

• Ano de defesa: 2020
• Autor(a) principal: D'Addio, Rafael Martins
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: Anotações de usuário; Information retrieval; Item representations; Recommender systems; Recuperação da informação; Representações de itens; Sistema de recomendação; User annotations
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/55/55134/tde-01122020-130836/

Recommender systems emerged as means of reducing the information overload problem. They require data to correctly function, i.e., users need to provide interactions so their profiles can be constructed, while items need descriptive data to be differentiated, especially in content-based approaches. Those approaches often rely on external sources, which can be structured metadata or unstructured texts such as synopses, news or user reviews. User reviews have increasingly been used as source of information due to their capability of conveying item characteristics as well as the authors opinion towards them. This doctorate research focuses on designing rich, semantic item representations from user reviews in order to increase recommendation accuracy. Beyond that, we aim to analyze their application in several domains, algorithms and recommendation tasks, verifying their quality and differences. We have proposed five conceptbased item representations, which are designed in the vector space model. They all share, in some level, the same feature set, varying in the semantics which their weighting scheme convey. We explore them in a plethora of approaches, algorithms and evaluation settings, from offline experiments to an online user trial. Results reveal the necessity of designing good, semantic item representations, which aid the recommender in correctly differentiating the items in the collection and ultimately producing more relevant suggestions to its users.