Poster Presentation at Boston Symmetry Day 2023

Many types of complex graphs (e.g., chemical molecules, biological molecules, knowledge graphs, etc.) can be naturally represented as attributed relational graphs (ARGs). The AGR representation extends ordinary graph representations by associating attributes/features with nodes and edges to characterize the corresponding entities and relationships, respectively, allowing the full representation of graphs with 3D geometry. ARGs are substantially more expressive, however, come with added complexities in training and analysis. Graph neural networks (GNNs) have been widely used in predicting the properties of molecular graphs. Existing GNNs do not effectively exploit ARGs and typically start with node attributes, which do not efficiently capture structural information. We hypothesize that the local structural context space can be well-characterized by a set of motifs, i.e., building blocks of chemical functional groups. Therefore, we propose a novel motif convolution module (MCM) for ARGs, allowing for better capture of local structural and semantic information from graph motifs.

Apr 7, 2023 1:00 PM — 2:30 PM