Duarte,Miguel Alexandre Marques2026-03-172026-03-172026http://hdl.handle.net/10400.5/117627Trabalho de projeto de mestrado, Engenharia Informática, 2026, Universidade de Lisboa, Faculdade de CiênciasSocial media has become one of the largest generators of data in recent years, and within this ecosystem, Instagram posts produce a wide range of public and private metrics that hold significant value for brands and stakeholders seeking to assess performance and engagement. In this project, state-of-the-art machine learning algorithms and techniques were leveraged, in order to improve current systems that did not make use of the available data and features. All types of content belonging to Instagram (Carousels, Images, Reels, Videos and Stories) were addressed, models were designed and tailored for each individual type of post, however, when data characteristics and feature behavior showed strong similarity, as in the case of Images and Carousels, an unified model was developed to avoid redundancy and improve generalization. All models predictions were compared against the capabilities of the current used algorithms, with the models outperforming them in every post demonstrating substantial improvements in overall stability of the predictions as well as reducing the error. For Stories however, performance was stunted by lack of feature diversity and the nature of available data. Given the dynamic nature of social media platforms: where algorithms, user behavior, and content trends evolve continuously monitoring, retraining, and model adaptation are essential. As new data is collected, further improvements can be achieved by integrating additional contextual features and optimizing model architectures, ensuring that the models remain accurate and relevant.application/pdfengMachine LearningSocial MediaRegressionTabular DataPrivate Instagram Metricsmaster thesis