Semantics-aware Dynamic Modeling Of Complex Social Systems

Complex social systems consist of two essential elements: semantics and social network.For social media with different forms like Weibo,Douyin,We Chat,their essence is that users interact with each other through semantic information based on the structure of the social network,generating rich behaviors under the dominance of semantics,thus promoting the evolution of the system.Since complex social systems are dynamic,dynamic modeling is an essential means to depict their dynamic characteristics.With the integration of social networks and content platforms in recent years,semantics play an increasingly significant role in social systems.Dynamic modeling of complex social systems with semantics has crucial scientific significance for explaining individual behavior patterns and revealing system evolution mechanisms.It also has broad application value in public opinion monitoring,churn prediction,content recommendation,etc.In this thesis,the machine learning method and statistical physics theory are combined to carry out the dynamic modeling of social systems from macro to micro levels with semantics.The dynamic models are useful in explaining the macro phenomenons and depicting micro mechanisms,as well as making accurate predictions,which provide both interpretability and accuracy.The main contributions and novelty of this thesis are as follows:· Dynamic modeling of information diffusion with semantics.Information diffusion is the core function of complex social systems,which drivesthe system evolution.Semantics results in the correlation of propagation behaviorsand show significant collective behaviors at the macro level.This thesis proposesmetrics to quantify collective behaviors.Through the novel latent semantic layerand temporal layer,this thesis successfully capture the collective behaviors and de-pict the dynamic characteristics of propagation.Finally,this thesis establishes anaccurate framework in various aspects of diffusion,including dynamics,structure,popularity,and semantics.This framework shows usefulness in many predictiontasks and has unique advantages in terms of dependence and accuracy.· Dynamic modeling of network structure evolution.The interaction between social networks and semantics is reflected in how the net-work structure evolves with semantics.Semantics introduce different types of edgesinto social networks,which changes the structure from homogeneous to heteroge-neous.This thesis finds that user churn in the network evolution presents a co-drivenmechanism of social and semantics.Thus,this thesis proposes the Social-Content-Mixture model to predict the churn probability over time.This model is useful forchurn prediction and recommendation and provides theoretical guidance for im-proving individual retention and intervening network evolution.· Dynamic modeling of individual activeness.The most intuitive expression of semantics in individual behavior is diversity re-flected by content and preference.Based on the systematical analysis on the dis-tribution and evolution of diversity,this thesis finds that the effects of diversity onactiveness have scenario heterogeneity and individual heterogeneity.Therefore,thisthesis proposes a dynamic model to quantify the effects of diversity on activeness.The parameters of this model have clear physical meaning and interpretability.Themodel is helpful to understand the individual behavior patterns,improve the designof social media,and enlighten controversial social issues such as filter bubbles andopinion polarization.