Aplicando técnicas de inteligência artificial e aprendizado de máquina para o monitoramento automático de publicidades de alimentos no Brasil

• Ano de defesa: 2023
• Autor(a) principal: Michele Bittencourt Rodrigues
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Minas Gerais; Brasil; ENF - DEPARTAMENTO DE NUTRIÇÃO; Programa de Pós-Graduação em Nutrição e Saúde; UFMG
• Palavras-chave em Português: Publicidade de alimentos; Alimentos industrializados; Televisão; Inteligência artificial; Aprendizado de máquina; Saúde Pública; Alimentos Industrializados; Inteligência Artificial; Aprendizado de Máquina; Dissertação Acadêmica
• Link de acesso: http://hdl.handle.net/1843/56062; https://orcid.org/0000-0001-7869-9665

Introduction: Food advertising has been recognized as one factor contributing to the choice of unhealthy foods and, consequently, to the increase in the prevalence of obesity and other chronic diseases related to food. Health organizations emphasize the importance of monitoring and restricting food advertising. Monitoring food advertising is crucial to assess the quality and quantity of food advertisements (ads) that are broadcast to the public, especially to children and young people. However, currently, the methods used for this monitoring are manual and of high execution complexity and costs, making the analysis time-consuming and costly. Consequently, the evidence produced is potentially limited in volume and quality, which may impair its translation into public policy. Objective: To develop a method based on artificial intelligence (AI) and machine learning techniques capable of automatically identifying and classifying food and non-food ad videos. Methods: A methodological study, which followed the protocol developed by the INFORMAS network (International Network for Food and Obesity/non-communicable Diseases Research, Monitoring and Action Support) for collecting data from food ads on television (TV). Food advertising data were collected from the programming of three free-to-air channels and two pay-for-view TV channels in Brazil in 2018, 2019, and 2020. The study included six execution stages: data collection, selection of training data, preprocessing, division of the initial database, training, and validation of computational models, and analysis of results. Results: The initial database was created from 2,124 hours of recording of Brazilian television programming and contained 703 food ads and more than 20,000 non-food ads. The results showed that the EfficientNet neural network associated with the balanced batches strategy achieved an overall accuracy of 90.5% on the test database, which represents a reduction of 99.9% of the time spent on identifying and classifying ads. Conclusion: The automatic identification and classification of TV ads into food and non-food ads using AI technology represents a promising approach with enormous potential to contribute to monitoring the food environment. This finding has important practical implications for researchers, public health policymakers, and regulatory bodies, not only in Brazil but also in other countries.