Evaluación y aplicación de diferentes estructuras de redes neuronales para grafos : evaluación comparativa de arquitecturas GNN.

Abstract

This study focuses on the development and comparison of hybrid Graph Neural Network models combined with Recurrent Neural Networks for urban traffic prediction, evaluating their ability to capture spatial and temporal dependencies. The ability to predict traffic is essential for governments and private companies to implement measures such as route planning, public transport expansion, and the improvement of navigation services. Using the METR-LA dataset, data preprocessing was carried out to normalize traffic characteristics and structure the urban road network as a graph. In this graph, the nodes represent traffic sensors, which may be located at intersections of two streets, and the edges represent the roads connecting these sensor locations. Traffic characteristics such as speed, flow, and occupancy were represented as node attributes. In this case, speed was used as the main feature to predict traffic. Several hybrid GNN models were implemented, including GCN-GRU, GCN-LSTM, GAT-GRU, and GAT-LSTM, considering both spatial and temporal dimensions. These models were trained and validated using crossvalidation techniques to ensure accuracy and avoid overfitting. The results were evaluated using standard metrics such as Mean Squared Error, Mean Absolute Error and Root Mean Squared Error.