Evaluación y aplicación de diferentes estructuras de redes neuronales para grafos : implementación de ensambles de modelos GNN.

Abstract

This study focuses on the development and comparison of ensemble models of graph neural networks (GNN) for urban traffic prediction, evaluating their ability to capture spatial and temporal dependencies. Accurate traffic prediction is crucial for governments and private companies to implement measures such as route planning, public transport expansion, and navigation service improvements. For this purpose, a widely recognized traffic dataset was used, in which the road network was modeled as a graph. Each node in the graph represents an intersection or road segment and contains attributes such as speed, detected through sensors installed along the roads. Several GNN models were implemented, including convolutional networks and attention networks, which, when combined with models that allow the study of sequential data, such as recurrent networks, better leverage both the spatial and temporal characteristics of the data. Additionally, ensemble strategies for these models were explored to improve prediction accuracy, employing techniques such as gradient boosting and averaging predictions from different base models. The results were evaluated using metrics such as Mean Squared Error (MSE) and Mean Absolute Error (MAE), showing that ensemble models provide significant improvements in the predictive power of graph neural network models.