MACHINE LEARNING FOR TEMPORAL HETEROGENEOUS GRAPHS: PREDICTIVE METHODS, INTERPRETABILITY AND APPLICATIONS.

This thesis develops a coherent framework for designing, evaluating, and interpreting predictive methods for temporal heterogeneous graphs by integrating modeling and tools from network science and graph deep learning. We introduce discrete-time graph learning architectures based on Graph Neural Networks (GNNs) and linear scoring functions tailored for forecasting dynamic, multi-relational networks, and provide a principled extension of message-passing paradigms to this setting. To address the limited interpretability of temporal graph models, we propose leveraging tools from temporal network analysis, such as link prediction heuristics, and we systematically benchmark explainability techniques in evolving relational contexts. Finally, we contribute novel high-resolution datasets derived from Web3 social platforms, enabling several applications for temporal graph learning in this context, such as link recommendation with textual content, transaction predictions, and user migration analysis. Together, these contributions advance both the theoretical foundations and practical applications of machine learning on temporal heterogeneous networks.