Motif Learning in Graph Neural Networks and Interpretation

Graph Neural Networks (GNNs) are widely used for modeling graph-structured data such as molecules, yet optimizing their representation learning capabilities and interpreting their complex predictive behaviors remain significant challenges. This dissertation addresses these problems by leveraging network motifs, which are recurring subgraphs that are closely associated with a graph's underlying properties. To enhance representation learning, we introduce the Heterogeneous Motif Graph Neural Network (HMGNN), which constructs a heterogeneous graph to learn atom-level and motif-level embeddings simultaneously. To improve interpretability, we propose MAGE, a generation-based explainer that uses motifs as fundamental elements to synthesize valid model-level explanations for trained GNNs, and MotifExplainer, which provides intuitive, motif-based instance-level explanations for individual graph predictions, rather than relying on simple nodes or edges. Finally, because the success of these methods depends heavily on the quality of the motifs themselves, we present MotifPiece, a data-driven approach that iteratively extracts highly effective motif vocabularies directly from molecular data. Together, these contributions demonstrate that improved motif construction and utilization are crucial for advancing both graph learning and interpretation.

Committee: Hongyang Gao (major professor), Qi Li, Wei Le, Ali Jannesari and Wensheng Zhang