Framework para sistemas de recomendação baseados em neural contextual Bandits com restrição de justiça

• Autor(a) principal: Santana, Marlesson Rodrigues Oliveira de
• Data de Publicação: 2024
• Tipo de documento: Tese
• Idioma: por
• Título da fonte: Repositório Institucional da UFG
• dARK ID: ark:/38995/001300000fprk
• Texto Completo: http://repositorio.bc.ufg.br/tede/handle/tede/13642
• topic: recomendação multistakeholder; justiça na recomendação; aprendizado por reforço; multistakeholder recommendation; fairness; reinforcement learning; CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO::SISTEMAS DE COMPUTACAO

The advent of digital businesses such as marketplaces, in which a company mediates a commercial transaction between different actors, presents challenges to recommendation systems as it is a multi-stakeholder scenario. In this scenario, the recommendation must meet conflicting objectives between the parties, such as relevance versus exposure, for example. State-of-the-art models that address the problem in a supervised way not only assume that the recommendation is a stationary problem, but are also user-centered, which leads to long-term system degradation. This thesis focuses on modeling the recommendation system as a reinforcement learning problem, through a Markovian decision-making process with uncertainty where it is possible to model the different interests of stakeholders in an environment with fairness constraints. The main challenges are the need for real interactions between stakeholders and the recommendation system in a continuous cycle of events that enables the scenario for online learning. For the development of this work, we present a model proposal, based on Neural Contextual Bandits with fairness constrain for multi-stakeholder scenarios. As results, we present the construction of MARS-Gym, a framework for modeling, training and evaluating recommendation systems based on reinforcement learning, and the development of different recommendation policies with fairness control adaptable to Neural models. Contextual Bandits, which led to an increase in fairness metrics for all scenarios presented while controlling the reduction in relevance metrics.