Research And Applications Of Uncertainty Propagation In Cascade Networks

With the rise and development of online social media,users are able to quickly generate online content,greatly enriching real-time communication among individuals.However,due to the uncertainty of the information propagation,understanding and analyzing information propagation patterns in social network has become a prominent research in both academia and industry.Once users receive information content on social platforms,they further propagate it on the network through retweeting behaviors,thus producing the phenomenon of information cascades.In the field of information dissemination,the information cascade is an important diffusion process,and accurately mining the evolutionary patterns of information cascades from complex and dynamical social networks has become one of the main challenges in cascade prediction tasks.In recent years,many research works have witnessed significant research efforts in information cascade prediction,yielding noteworthy achievements.Nevertheless,these works still faces with a number of urgent challenges.First,information cascades often exhibit irregularity in terms of their existence time,and common sequential models are inadequate in modeling the impact between such irregular events.Second,the uncertainty of information diffusion caused by the randomness of user behavior presents a major obstacle.Third,it is difficult to model the diverse social dependence relationships among users.Fourth,the complex time effects among users are frequently overlooked.Finally,the static learning paradigm of existing models limits their expressive power.To address challenges(1-2),this thesis proposes a novel information cascade model,Cas DO,which combines a probabilistic diffusion model with neural ordinary differential equations(ODEs).To more effectively capture real-world information dissemination processes,Cas DO incorporates a time-aware module that extends the discrete state transitions of traditional sequential models to continuous-time dynamics.The Cas DO model utilizes the probabilistic diffusion model to reconstruct the structural embeddings of users,capturing implicit relations among users in the social network.Cas DO also employs implicit neural ODEs to construct a time-varying generation process based on the initial implicit state of cascades,and capture the uncertainty in the time evolution process.To tackle challenges(3-5),this thesis introduces a novel meta-knowledge-driven information cascade framework,Meta Cas,which leverages adaptive meta-knowledge learned from user social structure,user preferences,and cascade temporal attributes to model information diffusion.Each information cascade prediction is treated as an individual task,which is transformed into a new adaptive task.The Meta Cas model incorporates a metaknowledge learning module,including meta graph attention network and meta long shortterm memory network,which extract cascade structural representations and dynamic cascade states while capturing the Spatio-temporal dependencies of different users in different information cascades through dynamically adaptive parameterization.This thesis conducts experiments on multiple large-scale cascade datasets,and the experimental results show that cascade frameworks surpasses existing baseline models.This study not only improves the accuracy of cascade prediction from the algorithm optimization level,but also quantifies the uncertainty of cascade propagation from the model reliability level.