COMBINING AND COMPARING HIERARCHICAL ATTENTION AND COMPOSITION-BASED GNNS FOR KNOWLEDGE GRAPH COMPLETION

Abstract

Graph Convolutional Networks (GCNs) have enabled the application of deep learning methods to large graphs. These models create an embedding representation for each node in the graph, and we train the model on these embeddings. The trained model can then be used to predict links between nodes or classify them. Link prediction, for instance, can be applied to biomedical graphs for tasks such as drug repurposing. By improving the performance of GCNs, we can enhance their application in drug repurposing. In this thesis, we aim to improve GCN results by enriching the representation of each node in the graph using two-hop paths for each relation.