Stat In-Person Colloquium: Dr. Cuneyt Akcora

Title: Topological Aspects and Neural Scaling Laws for Temporal Graph Foundation Models

Abstract: In recent years, there has been a surge in the development of foundational models for graph analysis that are capable of generalizing across various tasks and datasets. Temporal foundation models expand this concept into the temporal domain, providing essential tools for understanding and predicting patterns in evolving networks. These networks span diverse applications, from transaction networks to social dynamics.

In this seminar, we introduce a topological approach that enhances traditional methods in terms of both accuracy and efficiency. This method integrates topological data analysis with machine learning techniques to improve predictive accuracy while reducing computational demands.

We also present a foundational model that exhibits notable scaling properties: as the amount of pre-training data increases, the model's effectiveness on new networks also improves. This demonstrates the model’s capacity to scale and adapt, thus enhancing its predictive power.