Complex Systems can be modeled as collections of interactive agents exhibiting non-trivial collective behavior. The rich phenomena that occur in complex systems have intrigued researchers for a long time, yet modeling system dynamics and inferring interactions between agents remain difficult for conventional process-driven approaches.
Networks provide a simple model to understand and predict the emergent behavior of complex systems made up of a large number of interacting nonlinear dynamical units. One of the challenging and useful concerns in network science remains how structural properties of the underlying interaction network govern the dynamical behavior of interactions on the network, and how these rules can be discovered from real-world data.
With the advent of Artificial Intelligence, data-driven techniques, mainly based on deep learning, have become ubiquitous. For example, specialized artificial neural networks, called Graph Neural Networks (GNNs), are particularly tailored to model dependencies between linked agents on a graph. In fact, a complex system can be naturally represented as a dynamically changing graph. In this framework, modeling the system evolution corresponds to the problem of inferring the effect on an agent behavior of behavioral changes of neighboring agents or, in other words, estimating the state transition function at each node looking at the states of neighboring nodes. What remains hard to interpret are the characteristics of the interactions due to the well-known black-box nature of neural networks. However, some recent architectures, such as Graph Attention Networks, have traced a path toward interpretable GNNs by assigning attention to important neighbors, though a physical meaning is still hardly attachable to their outcomes.
The scope of this event is therefore to push the research on the cutting-edge problem of applying neural network techniques to the interpretation of complex systems’ dynamics, with a particular focus on real world applications. Specifically, we intend to stimulate the discussion on several actual and decisive topics, ranging from the evolution of physical ecosystems to the study of biological systems’ interactions (for example, in the human immune system, for genome-transcriptome-proteome data), from the analysis of communications to social networks and human behavior.
From a more theoretical point of view, we would like to discuss the recent developments in the mathematical theory and application of networks and networked dynamical processes, where the interactions among nodes cannot be decomposed into pairwise interactions. Applications of such approaches include analyzing persistent homology structure of data represented by higher-order networks, and uncovering new collective behavior and synchronization patterns that emerge from higher-order interactions. At the same time, multi-layer and modular networks have also found their applications in modeling biological and artificial neural networks and in addressing questions like how network structures affect collective behavior.
In this satellite, we are inviting researchers, many of whom have worked on both machine learning and complex systems. The goal is to bring together the scientists having expertise in traditional approaches in studying complex systems and networks and expert machine learning scientists for exchange of ideas, and formation of a platform for future collaboration, as well as to deliberate upon open problems in the complex systems and network science which can be addressed by machine learning techniques.