Analysis Of User Emotion Polarity Based On Multilevel Information Of Network

In recent years,with the rapid development of social media,the communication between network users is becoming more and more close,forming a huge and intensive social network.Users are no longer limited to sharing content with each other,and even express their emotion publicly.The emotion will spread along the whole social network,and then have an unknown impact on the user’s behavior.Therefore,sentiment analysis of social networks has attracted more and more attention.After studying the existing literature on the analysis of user’s emotional polarity,the author finds that most of the researches start from the user’s text data,that is,identifying the emotional tendency of the text through natural language processing technology.For the articles that study the user’s emotional polarity from the perspective of network structure,they are either based on simple friendship relationship or based on network embedding model,but these studies often have the defects of poor classification accuracy or poor interpretability.Therefore,based on the initial idea of improving the above shortcomings,this paper proposes a user sentiment analysis method by integrating multi-level network information.In this paper,we will use a large-scale data set about the polarity of users’ emotions to prove that there is a certain relationship between users’ emotions and their social relations from four different network levels.At the same time,we will prove that by fusing the implicit influence relationship between users in multi-level network,we can achieve higher accuracy results in sentiment analysis.Our main research work is as follows: 1.For two tuple networks,we mainly rely on the reciprocity theory(or homogeneity principle)in emotional networks.It is proved that only positive users show homogeneity in emotional networks,while negative users do not tend to establish contact with negative users as expected.2.According to the structure balance theory of emotional network,we extend the structure balance theory of social network to network nodes,and finally put forward the emotion balance theory.It is proved that in terms of emotional network,active users are more inclined to establish groups with other active users,and with the increase of negative members,the possibility of eventually forming triples will decrease.3.For the community network,we mainly study the influence of community size and community structure characteristics on users’ emotional polarity.It can be found that the smaller the scale of the community,the more similar the emotions of the community members.When the community connectivity is low,that is,when the community is separated from other parts of the network,the Subjective Well-Being(SWB)of the community is also low,and the community members are more inclined to express negative emotions.4.For the whole network,we prove that the real emotional network usually shows a core peripheral structure,that is,the network is composed of dense core and sparse peripheral,and active users are the core of the whole network,while negative users are the periphery of the network.In addition,we also improved the node2 vec algorithm based on the emotional relationship model.We set the weight of the edge to the intensity value of the emotional connection between nodes,and applied the weight to the sampling strategy,and finally obtained the low-dimensional vector representation of the nodes.5.Based on the network research of the above different levels,we will combine the features from different network levels,and select 41 features as input nodes by feature selection technology,and then use graph convolution model to classify the user emotion polarity.Finally,we have got more accurate prediction results.We use the feature information extracted from different levels of the network as the attribute feature of the node.One is to overcome the instability of the results caused by randomly generating node attribute features,the other is to make up for the shortcomings of single network level features by the fusion of multiple network level features,and third,to provide interpretability for the final model results.