Research On Temporal Cascade Evolution Trend In Social Networks

With the rapid development of the Internet and mobile networks,online social platforms have become closely intertwined with people’s daily lives.Social network users frequently interact on online platforms,resulting in a large amount of information content being spread on social networks.The process of information diffusion is called the cascade,which reflects the user interaction behaviors’ structural and temporal properties.Cascades often evolve over time and have a temporal structure as a whole,hence being called temporal cascades.Modeling and predicting the temporal cascade evolution trends in social networks is of great significance for understanding the information diffusion process and revealing diffusion laws.Thesis explains the temporal cascade evolution trends by exploring two tasks:predicting the information popularity and predicting the information burst time window.It summarizes the problems in existing work and provides corresponding solutions.Specifically:Current research on popularity prediction tasks mostly uses sequence sampling and subgraph partitioning to process the structural and temporal features of the cascade.They ignore the explicit time information in the cascade and cannot fully model the cascade role information.To address these issues,thesis proposes an explicit Time Embedding based Cascade Attention Network(TCAN)to handle the popularity prediction task.TCAN first uses a general time embedding method to embed explicit time information into cascade nodes,and then feeds the extracted two important role information(i.e.,cascade graph and cascade sequence)into the designed cascade graph attention network(CGAT)and cascade sequence attention network(CAST),respectively,to capture the temporal information of node interactions and user behaviors.Experiments on Weibo and APS demonstrate that TCAN has superior prediction performance and good interpretability.Thesis further captures the cascade properties by considering the diverse cascade relationships,while the cascade properties modeled by previous methods and TCAN can be explained by diverse cascade relationships,but the cascade relationships they capture are not complete.Therefore,thesis first explores the global spatio-temporal positional relationships,relative spatio-temporal interaction relationships,and interpersonal influence relationships between any two cascade nodes,and then proposes a Cascade-tailored Transformer framework(CasTformer)to predict information popularity.CasTformer uses a global spatiotemporal positional encoding(STPE)and relative relationship biases(RRB)to respectively learn the global spatio-temporal positional relationships and relative spatio-temporal as well as interpersonal interaction relationships between cascade nodes.It also uses a self-knowledge distillation technology to enable the model to learn better attention scores and more distinctive cascade representations to improve prediction performance.Evaluation experiments on Weibo,Twitter,APS,and DBLP datasets show that CasTformer can learn effective attention scores and cascade representations.Finally,thesis combines experience related to information popularity prediction to define the future burst time window prediction problem,and further proposes an end-to-end Deep Time Window Classification framework(DTWC)to predict the burst position.DTWC borrows encoders and decoders from the Transformer framework to encode the temporal features of time window popularity and sub-cascade graphs to generate the popularity values and burst window category probabilities on future windows.Experimental results on the Weibo dataset demonstrate that DTWC can learn the changing trends of the cascade in the future,thereby making accurate predictions.