How did the universe take shape after the Big Bang? What are the origins and evolutionary paths of life or species? How do brain networks evolve through plasticity to form intelligence? And how have social groups transformed over the course of history to become what they are today?
Answers to these intricate evolutionary questions might just be within reach. Recently, Associate Professor Hu Yanqing's research team from the Department of Statistics and Data Science at the Southern University of Science and Technology has developed a machine learning approach capable of reconstructing the historical formation of networked complex systems with high precision. This method offers a new research avenue for understanding the evolution of social systems, biological and ecological systems, among others. Their research paper has been published in Nature Communications and selected as an editor's highlight.
Complex networks, as a common form of networked complex systems, are widely present in fields such as biology, ecology, and social sciences. These networks primarily depict the interaction relationships among elements within complex systems and govern the system's functionalities. Exploring and understanding the evolutionary processes of complex networks is central to many fields, such as the origins and evolution of life and ecosystems, the reorganization of brain neural networks leading to intelligence, and the formation of communities and nations. However, complex networks often exhibit rich mesoscale structures, which simultaneously coexist within a network and interact in complex ways, collectively realizing the system's essential functions. Unveiling the relationships between these internal structures is often challenging, and further elucidating the evolutionary mechanisms of networks is even more difficult.
"In this study, we propose a machine learning approach to accurately reconstruct the historical process of each edge formation within a network based on structural information," said Dr. Zhang Yijiao, the co-first author of the paper, in an interview with the Chinese Science Bulletin. "This opens up new pathways for understanding the relationship between the structure and functionality of complex network systems, and provides important new research tools for various applications." She added, "By observing the current state of a system, we aim to reconstruct the evolution process from its initial state to the present. These problems are mostly fundamental issues in the scientific domain. If we can successfully reconstruct the generation process of each edge in a complex network, it implies that we can use a unified approach to reconstruct the detailed evolutionary processes of complex systems represented by networks. Undoubtedly, this will play a crucial role in deeply and meticulously understanding the physical laws of these systems and promoting the development of related application technologies."
In their research, the team demonstrated that, under certain conditions, machine learning techniques can accurately reconstruct the historical evolution trajectories of various complex networks, including protein-protein interaction networks, world trade networks, and social networks. Moreover, they confirmed the potential scientific value of the reconstructed evolutionary processes. For instance, the formation process of reconstructed protein-protein interaction networks can reveal some trends in the evolution of organisms at the molecular level. Integrating the obtained time-series information into network structure prediction methods can significantly improve prediction accuracy, holding potential application value in AI-driven drug discovery. Additionally, the reconstructed results can capture key structural features and mechanisms in the network formation process, such as preferential attachment, community structure, local clustering, and degree-degree correlation, which were previously unexplained by theory.
Diagram illustrating the process of network evolution and restoration methods (Image provided by the interviewee)
Reconstruction results help elucidate the evolutionary process of protein-protein interaction networks. (Image provided by the interviewee)
Improving Link Prediction Performance through Restoration Results (image provided by respondents)
Additionally, the team found through theoretical analysis that for large-scale networks, as long as the performance of the machine learning model is slightly better than random guessing, it can reliably reconstruct the entire historical formation process of the network. Since real complex networks are usually large in size, this conclusion suggests that the large-scale restoration of actual network evolution processes is typically highly reliable.
This research was supported by the National Natural Science Foundation.
Related paper information: https://doi.org/10.1038/s41467
