New PNAS Study on Predictability of Complex Networks

Led by Professor Qingpeng ZHANG, Associate Professor jointly affiliated with the HKU Musketeers Foundation Institute of Data Science (IDS) and the LKS Faculty of Medicine, a research team has developed a new theoretical framework to understand the predictability of complex networks. The study, led by Dr Fei JING, a Postdoctoral Fellow supervised by Prof ZHANG at HKU IDS, has been published in the Proceedings of the National Academy of Sciences (PNAS).

Professor Zi-Ke ZHANG of Zhejiang University and Professor Giorgio PARISI of Sapienza Università di Roma, Nobel Laureate in Physics (2021), are the corresponding authors of the study.

Complex systems underpin many aspects of the modern world, from artificial intelligence (AI) models to biological and social networks. Despite their importance, a fundamental question has remained unresolved: to what extent are such systems inherently predictable?

To address this, the research team introduced a rigorous theoretical framework by drawing on concepts from statistical physics. Specifically, the study maps the problem of network predictability onto the classical “spin glass” model, enabling a new way to analyse how connections in complex systems can be inferred.

A key breakthrough of the work is the demonstration that the global predictability of large networks can be decomposed into local contributions from individual connections. This insight significantly reduces computational complexity and provides a foundation for more efficient algorithms in analysing large-scale networks.

Building on this theoretical advance, the team proposed a highly efficient local sampling algorithm that relies only on neighbourhood-level information, substantially improving the scalability of network prediction methods.

The findings have broad implications across disciplines. In AI, the framework offers new ways to evaluate and design model architectures with improved efficiency and interpretability. In biomedicine, it may accelerate the prediction of molecular interactions, contributing to faster drug discovery processes.

This research highlights the value of interdisciplinary approaches that bridge statistical physics and data science, advancing our ability to understand and design increasingly complex systems.