Rank-based self-training for graph convolutional networks

Graph Convolutional Networks (GCNs) have been established as a fundamental approach for representation learning on graphs, based convolution operations non-Euclidean domain, defined by graph-structured data. GCNs and variants achieved state-of-the-art results classification tasks, especially in semi-supervised scenarios. A central challenge consists how to exploit the maximum of useful information encoded unlabeled In this paper, we address issue through novel self-training improving accuracy tasks. margin score is used rank-based model identify most confident sample predictions. Such predictions are exploited an expanded labeled set second-stage training step. Our suitable different GCN models. Moreover, also propose rank aggregation sets obtained The experimental evaluation considers four variations traditional benchmarks extensively literature. Significant gains were all evaluated models, reaching comparable or superior state-of-the-art. best combinations