Waiting times have been identified as a significant problem within the public transportation system, leading to a poor user experience. Emotions generated by this issue have been evaluated with artificial intelligence (AI) models to assess the impact on public transport performance. However, the relationship between waiting times and the actions that buses should take according to user emotions had not been considered. The main objective of this work was the design and implementation of a hybrid model based in real-time passenger feedback for public transportation, integrating a natural language processing (NLP) model and multi-agent RL (MARL). An NLP called RoBERTa was used for the analysis of passenger feedback, and an independent Q-learning method was generated to improve bus frequency based on the recorded emotions. The analysis revealed an accuracy of 98.2% in identifying emotions, which were categorized as bad, good, or neutral. Consequently, public transport frequencies were increased by up to 50%. Furthermore, a scheme oriented to graph neural networks (GNNs) was designed in GraphSAGE, where nodes were represented as routes and edges as the interactions with emotions; an accuracy of 94.36% was obtained. In conclusion, these findings suggested that the public transport system can be optimized according to user emotions and that diverse actions can be taken depending on the emotion. Future work will focus on creating an interconnection model, as it has been established that connected routes mutually affect each other. Local and global optimization would be defined, while MARL models would be preserved, coordinating routes and buses.