UCLA

Graph is a pervasive data type in the real world, as it serves as a succinct yet powerful abstraction for entities and their interconnections. Consequently, generating high-quality graph representations that encode the graph information is important, as these representations would be instrumental in graph-related tasks. In this context, Graph Neural Networks (GNNs) have emerged as a significant advancement and gained prominence for their ability to learn powerful graph representations which lead to state-of-the-art performance across a variety of graph-based applications.

However, despite their remarkable capabilities, the design and training processes of GNNs are still fraught with challenges. Given this, my research goal is to identify and address these hurdles from both the effectiveness and efficiency perspectives, aiming to develop GNN models that are potent yet scalable, thereby enhancing GNN’s power in producing superior graph representations in practice. This dissertation systematically investigates two fundamental questions: (1) What impedes the effectiveness of GNNs? (2) How to train GNNs efficiently? Despite providing a comprehensive discussion of each question, it presents practical solutions to mitigate each side and covers both homogeneous graphs and heterogeneous graphs.