Research On Information Evolution And Blocking Model Of Complex Dynamic Network

Complex dynamic networks are widely present in various real-world application scenarios,where information and even network structures on the network are constantly changing over time.Network data at different time periods correspond to different data generation distributions,which contain different temporal patterns.This poses a significant challenge to modeling the information evolution process on networks.Efficient modeling of the information evolution process on complex dynamic networks is of great significance to promote the development of related fields.Moreover,in typical complex dynamic networks,the social networks,due to their wide audience,low barriers to access,and ease of dissemination,they have become a breeding ground for the spread of malicious information such as rumors.The unfettered spread of malicious information such as rumors can undermine social stability and cause significant public loss.Therefore,constructing an effective network information blocking model is of great importance for intervening in the spread of malicious information such as rumors and maintaining social stability.This paper focuses on the research of information evolution and blocking models on complex dynamic networks,which mainly includes the following two aspects:Existing methods for modeling information evolution on dynamic networks typically utilize tools such as Recurrent Neural Networks(RNNs)to capture temporal patterns from historical data and use these patterns to infer the network’s future evolution,assuming that future network data contains the same temporal patterns as the historical data.However,the data generation distribution of a graph may change over time,introducing new temporal patterns in future periods that render the temporal patterns learned from historical data ineffective.To address this challenge,this paper proposes a model-agnostic framework entitled Dynamic Feature-wise Linear Modulation(DyFiLM),which uses a hypermodel at the task level to learn new temporal patterns from future input data.Specifically,the DyFiLM framework takes existing dynamic graph representation learning models as base models and modulates them using the hypermodel to express different temporal patterns on data from different periods.DyFiLM is capable of capturing newly emerging temporal patterns on unseen data,thereby enhancing the modeling capacity of information evolution on dynamic networks.This paper applies the proposed DyFiLM framework to three different base models and verifies its effectiveness on four datasets through experiments.The results demonstrate that compared to relevant methods,DyFiLM significantly improves the modeling performance.Currently,research on network information blocking models mainly relies on the modeling of information propagation in networks using the framework of propagation dynamics,where the effectiveness of blocking is measured by the change in the spread of competing information after introducing one’s seed nodes.However,user opinion is an important factor for measuring the impact of information propagation,and ignoring its influence by information spread may limit the effectiveness of blocking models.Therefore,this paper proposes a competitive linear threshold model with opinion dynamics to model the impact of information flow on user opinion.This model integrates opinion dynamics into the propagation dynamics model and measures the impact of information spread from the perspective of user opinion.Then we construct an equivalent two-stage model that decouples the complex information propagation and opinion exchange processes.We define the information blocking optimization objective from the perspective of user opinion based on the second-order model,and prove its monotonicity and submodularity to guarantee the effectiveness of the greedy algorithm used for approximation.Finally,we propose the CLDAGWOD algorithm from the perspective of local approximation to quickly simulate the propagation process for each round of the greedy algorithm.The performance of the proposed CLDAGWOD algorithm is experimentally validated on three real social network datasets,and the results show that it effectively achieves blocking.