Learning on Graphs with Robustness, Efficiency and Adaptability

Graphs are ubiquitous data structures that capture complex relationships among entities in real-world systems, including social networks, biological networks, transportation systems, and e-commerce platforms. With the rapid growth of graph data, numerous challenges and tasks have emerged, significantly impacting everyday life. While Graph Neural Networks (GNNs) have become the dominant approach for addressing these challenges, they face critical limitations in robustness, efficiency, and adaptability that restrict their broader application. This thesis systematically addresses these limitations to develop more powerful GNN models. To improve robustness, this thesis investigates GNNs from both model and data perspectives. On the model side, it identifies dataset shift as a fundamental issue and proposes a corresponding solution. On the data side, it introduces a data augmentation method to generate predictive and concise graph representations. To enhance efficiency, the thesis presents techniques that improve GNN expressiveness through efficient graph structural estimation. Additionally, it proposes a novel training paradigm that bypasses the conventional gradient descent process, allowing GNNs to fit directly to data. To improve adaptability, the thesis explores methods for GNNs to generalize to new data and tasks. Specifically, it introduces a framework for link prediction that can adapt to arbitrary graphs during inference. Furthermore, it develops a practical application of GNNs for predictive tasks in relational databases. By addressing these limitations, this thesis advances the robustness, efficiency, and adaptability of GNNs, expanding their applicability in diverse real-world scenarios.