UNRI · Unifying Neural Relational Inference

Our world, from the human brain to the global economy, is built on complex networks. Understanding the structure of these networks is a cornerstone of modern science and technology, yet these connections are often difficult to spot. Typically, we only observe a system's dynamic activity over time, such as the fluctuating signals from brain regions or the daily prices of stocks. This gives rise to a critical scientific question: can we create an accurate map of a system's hidden and evolving connections (its latent dynamic relationships) simply by watching the activity of its components? This computational challenge boils down to understanding the underlying rules and relationships in a system, represented by a graph, just by observing how its parts evolve over time. In the field of machine learning, this is known as Neural Relational Inference (NRI). Existing approaches are designed for either static or dynamic graphs, and are almost exclusively transductive, which in the NRI context means being optimized on a single dataset. To address this gap, my project, UNRI, will pioneer a novel class of temporally consistent diffusion models. The core idea is to learn the ""rules of change"" by modeling the generative process of the graph's evolution itself. My model will treat the structure at each timestep as a direct, stochastic evolution from the previous state, ensuring temporal consistency by design. My project will deliver three main contributions. The first is a concrete, high-performance model that solves a pressing technical challenge by operating across all key NRI settings.Inspired by my supervisor's successful work on unifying a particular field of graph machine learning, as a second contribution, I will develop a general framework for the entire NRI field and use it as the foundation to build a theoretically-grounded open-source library.""