Graph-Based Machine Learning for Real-World Applications

Abstract

Graph-based machine learning has emerged as a powerful approach for modeling complex real-world data, particularly in finance and medical research, where data are heterogeneous, noisy, incomplete, and highly structured. This thesis investigates graph-based machine learning methods for addressing practical challenges in financial and medical AI, with an emphasis on robustness, interpretability, and fairness.

The first part of the thesis focuses on financial applications. A heterogeneous graph learning framework is developed for credit card fraud detection by modeling complex transactional relationships. Feature importance–based edge weighting is incorporated into a graph neural network to improve predictive performance and model interpretability. The thesis further investigates fairness in graph neural networks by proposing a framework that balances group fairness and individual fairness, demonstrating that these objectives can be jointly improved in graph learning tasks.

The second part of the thesis focuses on medical AI for retinal disease research. Machine learning methods are applied to predict subretinal lesion severity and identify genes associated with disease progression from high-dimensional transcriptomic data. Building on this work, a graph pseudotime analysis framework is proposed to infer disease progression trajectories, while neural stochastic differential equations are employed to characterize pathway dynamics and identify transition points.

Overall, this thesis demonstrates that graph-based machine learning provides a flexible and effective framework for modeling complex data across financial and biomedical applications. The proposed methods improve predictive performance, interpretability, and fairness while providing biologically meaningful insights into disease progression.