Link prediction with Simple Graph Convolution and regularized Simple Graph Convolution

Abstract

Attributed graphs are used to model real-life systems in many domains such as social science, biology, etc. Link prediction is an important task on attributed graph with a wide range of useful applications. Simple link prediction approaches have limitation in their capability to capture network topology and node attributes. Graph Neural Networks (GNNs) provide an efficient framework incorporating node attributes and connectivity to produce informative embeddings for many downstream tasks including link prediction. In this work, we study two variants of GNNs, namely Simple Graph Convolution (SGC) and its extension for link prediction on three citation datasets. While it is fast and efficient, our model is insufficient to capture the complex node connectivities. On the other hand, imposing regularization reduces overfitting and improves model performance.