R-GCN: The R Could Stand for Random
This challenges the necessity of learned parameters in graph neural networks for knowledge graphs, potentially simplifying model design and training, though it is incremental as it builds on existing R-GCN frameworks.
The paper argues that the main contribution of Relational Graph Convolutional Networks (R-GCNs) is the message-passing paradigm rather than learned weights, and introduces Random R-GCNs (RR-GCNs) with no trained parameters, showing they can compete with fully trained R-GCNs in node classification and link prediction tasks.
The inception of the Relational Graph Convolutional Network (R-GCN) marked a milestone in the Semantic Web domain as a widely cited method that generalises end-to-end hierarchical representation learning to Knowledge Graphs (KGs). R-GCNs generate representations for nodes of interest by repeatedly aggregating parameterised, relation-specific transformations of their neighbours. However, in this paper, we argue that the the R-GCN's main contribution lies in this "message passing" paradigm, rather than the learned weights. To this end, we introduce the "Random Relational Graph Convolutional Network" (RR-GCN), which leaves all parameters untrained and thus constructs node embeddings by aggregating randomly transformed random representations from neighbours, i.e., with no learned parameters. We empirically show that RR-GCNs can compete with fully trained R-GCNs in both node classification and link prediction settings.