Data-Driven Dynamical Modeling Of Complex Social Systems

Complex social systems,such as WeChat,Weibo,Facebook,Twitter,Tik Tok,etc.,accept users' input signals,generate output signals,meet users' various information and social media needs,constantly evolve over time,and generate social big data.Explaining the evolutionary mechanisms of complex social systems is the core topic of this paper.It is of great significance to understand the scientific problems of how complex systems evolve in nature,and how to provide interpretable recommendation,advertising and other computing-based services in social systems.However,to study the evolutionary mechanisms of complex social systems is extremely challenging.First,the complex social system is composed of hundreds of millions of linked individuals,and its output signals are always in the form of networked data,exhibiting structural complexity.Second,microindividuals in complex social systems interact with each other dynamically,resulting in a large difference between the macroscopic output and the sum of micro-individual inputs,that is,nonlinearity;or resulting in large-scale emergence on a short time scale,that is,burst;these two are collectively called dynamic complexity.The complex social system presents a random disorder state at the microscopic level,but the macroscopic phenomena determine the ordered state,and thus its evolutionary process exhibits multiscale complexity.Traditional analysis methods are based on physical dynamical models,trying to characterize the dynamics of complex social systems and reveal the mechanisms of their changing.However,complex social systems in the real world such as WeChat,Weibo,etc.have billions of nodes and tens of billions of edges,and their dynamic phenomena are reflected in different scales and different scenarios such as macro network evolution,micro-individual social behavior,dynamic propagation of information on the network and so on.Through a data-driven approach,we have discovered many new complex phenomena of the above-mentioned complex social systems in different scales and different scenes,and the traditional dynamical models failed in capturing the observed phenomena.This paper tries to combine the theories in both computer science and physics to model the evolutionary mechanisms of complex social systems through a data-driven dynamical modeling approach.Specifically,this paper studies three core sub-topics of the evolution of complex social systems.First,the discovery and modeling of multi-scale evolution of social networks,which aims to answer the question of how complex social systems grow at different scales.Second,Complex pattern formation of information flow,which is designed to answer the question of how information flows over complex social networks.Third,the theorem of the dynamic origin of distribution functions,which is designed to answer questions about how to connect micro-behavior and macrophenomena.Our proposed methods have been experimentally verified on large-scale social network datasets such as WeChat(first time),Tencent Weibo and so on.For the discovery and modeling of multi-scale evolution of social networks,we find that the nodes and links of social networks follow the power-law growth over time,and we propose the NetTide model which accurately captures the evolutionary power law,and improves the performance of long-term network evolutionary prediction.Furthermore,we find long-term nonlinear growth and short-term bursty growth of micro-network evolution.We propose the long and short memory stochastic process,which accurately describes the dynamic random behavior of individuals and improves the predictive performance and interpretability of micro-social behaviors.For the complex pattern formation of information flow in the network,we find the complex structure of information flow in the real social systems,and then quantify their complex geometric patterns,and propose a data-driven heterogeneous branching processes to explain the mechanisms of complex pattern formation of information flow,which greatly improves the accuracy of fitting complex geometric structures.For the dynamic origins of distribution functions,we propose to explain that the distribution function is obtained from the randomly arriving microscopic individuals through a determined dynamic process,which connects the microrandomness and the macro-determinism.The theorem greatly improves the accuracy and interpretability of fitting complex multi-scale distributions.Through a data-driven dynamical modeling approach,we have well characterized the new observed phenomena in the complex social systems,and further explained their dynamic evolutionary mechanisms from the perspective of complex systems,which effectively combine the computability of data science and interpretability of statistical physics.Our proposed data-driven dynamical modeling approach attempts to lay a theoretical foundation for modeling,understanding,and predicting large-scale complex social systems in the real world.