Filtragem de sinais sobre grafos aplicada à classificação de dados ruidosos definidos em estruturas irregulares

Abstract

Graph convolutional networks (GCN) allow deep neural networks to learn from data defined in irregular structures represented by a graph. This approach has attracted more and more attention in recent years. Such algorithms generally learn based on the information associated with the vertices and edges present in the graph, being possible to use them to improve the performance of several tasks such as classification of vertices, a signal defined on the graph or the graph itself. The objective of this work is to evaluate the classification of noisy data arranged in an irregular structure from the training of low-pass filters performed by the convolutional layers of a GNN. Filters are designed and applied based on signal processing tools on graphs such as the graph Fourier transform. The graph-based convolutional neural network will perform the learning of parameters for different noise scales in the graph Fourier transform domain that are able to extract relevant information for classification of the noisy dataset.