From Chaos to Harmony: Addressing Data De-Noising, Complexity and Adaptability in Graph Machine Learning

Graph representation learning—especially via graph neural networks (GNNs)—has demonstrated considerable promise in modeling intricate interaction systems, such as social networks and molecular structures. However, the deployment of GNN-based frameworks in industrial settings remains challenging due to the inherent complexity and noise in real-world graph data. This dissertation systematically addresses these challenges by advancing novel methodologies to improve the comprehensiveness and robustness of graph representation learning, with a dual focus on resolving data complexity and denoising across diverse graph-learning scenarios. In addressing graph data denoising, we design auxiliary self-supervised optimization objectives that disentangle noisy topological structures and misinformation while preserving the representational sufficiency of critical graph features. These tasks operate synergistically with primary learning objectives to enhance robustness against data corruption. The efficacy of these techniques is demonstrated through their application to real-world opioid prescription time series data for predicting potential opioid over-prescription. To mitigate data complexity, the study investigates two complementary approaches: (1) multimodal fusion, which employs attentive integration of graph data with features from other modalities, and (2) hierarchical substructure mining, which extracts semantic patterns at multiple granularities to enhance model generalization in demanding contexts. Finally, the dissertation explores the adaptability of graph data in a range of practical applications, including E-commerce demand forecasting and recommendations, to further enhance prediction and reasoning capabilities.