Research On Cascade Prediction Of Information Based On Fusing Network Topology

The cascade prediction of information diffusion is based on the observed cascade sequence to predict the most likely infected nodes,which can be applied in scenarios such as public opinion dissemination control and false news detection.The dynamic method focuses on researches such as propagation mechanism,speed of spread and threshold of infection,but fails in the cascade prediction task.Although existing deep learning methods can improve the prediction results,they ignore the influence of network topology structure influence,which make the methods cannot explain the difference of prediction results in different cascades.This paper proposes the predictability of information dissemination cascade prediction fusion network topology,which is constructing a small-scale set of potential infections in large-scale networks.This paper proposed end-to-end deep learning frameworks by fusion the characteristics of first-order/second-order and community structure to learning the differences of cascades,which significantly improves cascade prediction performance.The methods integrates cascade timing characteristics,inter-user propagation influence and network homogeneity effectively.Thus,in view of the different scenarios of the predictability of cascade: community structure,the influence relationship between users,and the information dissemination under the influence of cascade size heterogeneity,the cascade prediction models are constructed respectively.The specific contributions are as follows:A new research proposal for the influence of the network structure on the cascade prediction of information dissemination is proposed.The paper analyze the influencing factors of cascade prediction from multiple angles such as information dissemination cascade characteristics,network topology,information,etc.,and visualize the community structure,first-order/second-order neighbor structure’s impacts for the predictability problem of cascade prediction.A cascading prediction model of information dissemination integrating community structure is proposed.Network topology homogeneity is the core factor that affects cascading prediction.Based on the community structure,the potential candidate set can be focused on the community scale,while solving the challenges brought by the sparsity of online social networks.Combined with the LSTM modeling cascading time series feature,a unified representation learning framework is constructed,and the prediction effect is better than baseline on multiple data sets.A cascade prediction model of information diffusion integrates the first-order and second-order neighbor structures of users is proposed,which greatly reduces the candidate set of potential infections in large-scale networks,and solves the problem caused by the local clustering of infection sequences in the cascade.The generation of cascading differences is driven by the first and second order topology,and the set of potential candidate infected nodes is uneven.The proposed method improve cascading prediction performance and enhance interpretability.On the basis of the second model,a cascade prediction model based on self-attention mechanism is proposed.The first and second order neighbors of nodes in the longsequence cascade almost cover the entire network,so that the size of the potential infection set cannot be reduced.Thus,this paper propose a framework for adaptively selecting infected nodes,and the self-attention mechanism and network embedding technology are used to alleviate the unpredictable problem of long sequence cascades.This paper solves many challenges brought by cascading differences.From the perspective of predictability,it incorporates the characteristics of the network topology,supplementing and perfecting the model of cascading prediction of information dissemination based on deep learning.