Dynamic Weight Prediction Based On Comprehensive Embedding Of Node Features In Signed Networks

Weight is an important indicator to measure the closeness of the relationship between nodes in the network.Predicting the weight in the network can help people analyze the behavior and activity trends in the social network.At present,most researches focus on the weight prediction of unsigned networks.But in real life,the links of social networks have signed information,and the network structure as well as weight will change over time.For weight prediction of dynamic signed networks,it is very difficult to predict the complex relationship between signs and weights as well as the nonlinearity of time-varying.A graph neural network learning model based on node comprehensive feature embedding is proposed to deal with the above challenges,which can realize the embedding of signed networks and the weight prediction of dynamic signed networks.Aiming at weight prediction in signed networks,this thesis puts forward a signed network embedding model based on graph neural network.The model integrates balance theory,status theory and weights in signed networks to learn node feature representations to accomplish weight prediction in signed networks.So as to learn node feature representation,this paper makes use of two different graph neural networks to obtain node weight feature representation and node sign feature representation respectively.Next,it utilizes attention mechanism to learn the different importance of features to fuse node feature representation.Then using fully connected neural network to implement weight prediction.Compared with other methods,the RMSE value of this model is reduced by 8.8% on average,and the PCC value is increased by 16%.For weight prediction in dynamic signed network,this article proposes a weight prediction model for dynamic signed network based on the comprehensive feature embedding of signed network nodes.The model utilizes a graph neural network to obtain the comprehensive feature representation of nodes at different time snapshots.Then the temporal self-attention mechanism is used to capture the evolution characteristics of each signed network node in the time series.In order to integrate the feature representations of nodes at different time snapshots,this thesis takes advantage of long shortterm memory network to embed node sequence features.Finally,it uses a fully connected neural network for dynamic weight prediction.Experiments show that the RMSE value and MAE value are reduced by 5.2% and8.9% on average compared with other methods,which proves that the model can improve the performance of dynamic signed network weight prediction.