Information Diffusion Modeling And Popularity Prediction Based On Neural Point Process

With the rapid development of information technology,sharing information through social media platforms such as Twitter,Facebook,and Weibo has become the main channel for individuals to express their opinions and views on various topics.The original information item posted by the user and the subsequent sharing and forwarding of other users constitute an information cascade.Understanding the information propagation mechanism,and modeling the information cascade diffusion process reasonably,is crucial to predict its popularity in recent social network research and directly impacts various downstream tasks,including rumor detection,public opinion analysis,etc.To model the information diffusion process,current research on information cascade popularity prediction primarily employs simple sequence methods or shallow graph neural networks.There are shortcomings and difficulties persist,despite the significant progress that has been made.The existing methods mainly apply simple sequential methods,such as recurrent neural networks,are difficult to effectively simulate irregular events during the process of information forwarding in the discrete time domain;their shallow graph neural networks cannot capture long-term user dependencies;Existing methods mainly model the topological structure and temporal characteristics of information diffusion separately,hence its coupled characteristics during the diffusion process cannot be simulated;In addition,there is still a lack of effective methods to embed information cascades into lowdimensional vector spaces.To address these challenges,this thesis proposes a Transformer-enhanced Hawkes process Hawkesformer,which combines a hierarchical attention architecture with a Hawkes self-exciting point process to model information cascades and use them for the final popularity prediction.Hawkesformer extends the traditional Hawkes process from the topological level and efficiently extracts information from the continuous time domain.Specifically,it links a two-level attention architecture to parameterize the strength function of Hawkesformer,the first-level simulates the coupled topological and temporal dependencies by distinguishing primary and nonprimary paths to capture the global dependencies between nodes in the graph;The secondlevel learns the cascade evolution rate by a local pooling attention modules for modeling short-term changes.Moreover,existing methods cannot cope with datasets that obey long-tail distributions;The majority of research focuses on model design for performance improvement,lack of data mining and utilization.In order to solve the challenges,this thesis proposes a novel method DeCas to deal with the long-tailed cascaded data,which decouples the model training into two stages: feature extraction and regressor fine-tuning,supplemented by multiple sampling strategies and regressor strategies to reduce the popularity prediction bias.In addition,DeCas,as a top-level general method,can be incorporated into any information cascade popularity prediction model,enabling it to deal with the long-tailed data.Finally,this thesis conducts extensive experiments and a series of ablation studies on multiple benchmark datasets,and verifies the effectiveness and rationality of Hawkesformer and DeCas through comparative analysis with existing state-of-the-art models,and also demonstrates the strong interpretability of the two models.