Missing Link Identification from Node Embeddings using Graph Auto Encoders and its Variants

Graph representation learning recently has proven their excellent competency in understanding large graphs and their inner engineering for various downstream tasks. Link completion is an important computational task to guess missing edges in a network. The traditional methods extract local, pairwise information based on specific proximity statistics that are always ineffective in inferring missing links from a global topological perspective. Graph Convolutional Network (GCN) based em-bedding methods may be an effective alternative. In this work, we try to experimentally assess the power of GCN-based graph embedding techniques, namely Graph Auto Encoder (GAE) and its variants GraphSAGE, and Graph Attention Network (GAT) for link prediction tasks. Experimental results show that the GAE-based encoding methods are able to achieve superior results for predicting missing links in various real large-scale networks in comparison to traditional link prediction methods. Interestingly, our results reveal that the above techniques successfully recreate the original network with high true positive and negative rates. However, it has been observed that they produce many extra edges with an overall very high false positive rate.