Análise de redes neurais baseadas em grafos

Abstract

The recent success of applications and the wide development in the deep learning field has obtained recognition in numerous machine learning tasks, such as computer vision and natural language processing. However, multiple learning tasks involve data that can't be represented in the euclidian domain and have complex structural relationships, such as graphs. Considering the complexity of adapting traditional deep learning techniques on graphs, Graph Neural Networks were introduced. Such networks are learning models capable of extracting representations of non-euclidian data, allowing the execution of learning tasks in this type of structure. Recently, different studies have achieved the state of the art in a wide variety of graph classification tasks. Therefore, the present work analyzes distinct architectures of Graph Neural Networks in the node, edge and graph classification tasks with representative datasets of social, biological and citation networks.