Towards Open World Graph Learning and Applications

Abstract

Graph learning has emerged as a foundational paradigm for modeling interconnected data across financial networks, industrial systems, and social platforms. However, the remarkable progress of graph neural networks has largely been built upon closed-world assumptions that fundamentally mischaracterize real-world applications: graphs evolve continuously across multiple timescales, exhibit heterogeneous structures that connect dissimilar entities through heterophilic relationships, contain critical yet rare events under severe label scarcity, and require sub-second processing under strict reliability constraints. In this dissertation, we propose a research agenda organized around three integrated modules for open-world graph learning: (M1) Theoretical Foundations---establishing rigorous frameworks that characterize learning dynamics on heterophilic graphs through spectral analysis and convergence guarantees; (M2) Algorithmic Innovations---developing universal architectures that simultaneously handle temporal dynamics, structural heterogeneity, and label scarcity; and (M3) Benchmarks and Applications---creating comprehensive evaluation infrastructure and validating the proposed frameworks across domains. Within this agenda, M1 explores and characterizes open-world graph patterns; M2 translates theoretical discoveries into practical systems that integrate multimodal signals and generate augmented representations; and M3 validates these frameworks through standardized benchmarks and real-world deployments while providing feedback to refine M1 and M2. Overall, this dissertation develops principled frameworks that advance graph learning from static homophilic graphs to evolving heterophilic networks, and from homogeneous structures to heterogeneous multimodal data under label-scarce and reliability-constrained settings.