Improving the Relational Inductive Bias of Graph Neural Networks
Supervisors
Suitable for
Abstract
Graph neural networks (GNNs) have achieved state-of-the-art results on a wide range of tasks e.g., node and graph classification, on various types of graphs, owing to their natural encoding of graph structure and properties. However, their potential remains yet to be unlocked for multi-relational data (i.e., graphs where edges can have multiple labels). In particular, while these models account for relation types during message passing (see, e.g., rGCN [1]), they typically do not incorporate any explicit relation-specific inductive bias, e.g., inherent relational properties such as symmetries, relational hierarchies. The goal of this project is to design a novel GNN model with relation-specific inductive bias and study its relational properties, both theoretically and empirically.
Prerequisites: Good programming skills, experience with Pytorch/Tensorflow
References: A classical base GNN model for handling relational data is rGCN [1], an extension of GCN [2] with relation-specific message passing.
[1] Schlichtkrull et al. Modeling Relational Data with Graph Convolutional Networks. ESWC 2018.
[2] Kipf and Welling. Semi-Supervised Classification with Graph Convolutional Networks. ICLR 2017.