Expanding Label Sets for Graph Convolutional Networks

• DOI: 10.48550/arxiv.1912.09575
• Publication Date: 2019-01-01
• Authors: Mustafa Coşkun; Burcu Bakır-Güngör; Mehmet Koyutürk
• Keywords: FOS: Computer and information sciences; Computer Science - Machine Learning; 0303 health sciences; 03 medical and health sciences; Statistics - Machine Learning; Machine Learning (stat.ML); Machine Learning (cs.LG)

In recent years, Graph Convolutional Networks (GCNs) and their variants have been widely utilized in learning tasks that involve graphs. These include recommendation systems, node classification, among many others. classification problem, the input is a graph which edges represent association between pairs of nodes, multi-dimensional feature vectors are associated with some nodes known labels. The objective to predict labels not labeled, using features, conjunction topology. While GCNs successfully applied this caveats they inherit from traditional deep models pose significant challenges broad utilization classification. One such caveat training GCN requires large number labeled instances, often case realistic settings. To remedy requirement, state-of-the-art methods leverage network diffusion-based approaches propagate across before GCNs. However, these ignore tendency diffusion biasing proximity centrality, resulting propagation well-connected graph. address here we present an alternate approach extrapolating following three steps: (i) clustering identify communities, (ii) use algorithms quantify each thereby obtaining low-dimensional topological profile for node, (iii) comparing profiles most similar nodes.