Research On Key Technologies Of Online Social Network Public Opinion Analysis For Public Opinion Elements

"Information Security Dictionary" defines public opinion as a collection of four elements: people’s cognition,attitude,emotion and behavioral tendency towards events.Most of the research on online social network public opinion analysis technology revolves around these four elements.Especially with the continuous development of social network technology in recent years,sufficient data and supporting models have been provided for the quantitative analysis of public opinion elements.Relevant research has achieved great progress.A lot of achievements have been made,but there are also many problems.For the cognitive element,the first is to solve the problem of authenticity.With the influx and dissemination of a large amount of false information on social networks,there are many examples of public opinion incidents caused by users’ cognitive biases towards false information.Although relying on rumor detection technology can distinguish the authenticity of information in a timely manner,the existing rumor detection technology is too dependent on the correctness of feature selection,and the ability to express the structural features of social networks is not strong,and most deep learning algorithms have high complexity and low efficiency.For emotion and attitude elements,a key issue is how to conduct quantitative analysis more accurately.Public opinion information often contains multiple aspects,so fine-grained sentiment analysis technology is required.Although the latest sentiment analysis technology can use the graph neural network to analyze the grammatical structure features of sentences and obtain the emotional polarity of users on different aspects of public opinion information,for the multi-layer grammatical structure,the graph neural network must be adapted by increasing the number of layers,resulting in problems such as high model complexity and high overhead.For the behavior tendency element,the core issue is to evaluate the possible impact of user behavior,which can be realized through the cascading prediction of public opinion information.Many current studies rely on deep learning to predict cascades end-to-end,but deep learning hidden layer features lack interpretability,and cannot intuitively and accurately describe the different effects of user interaction on cascades under different topics,which is not conducive to information dissemination analysis and management.People’s demand for public opinion analysis is more reflected in the evaluation of the importance of public opinion information based on quantitative analysis of various elements of public opinion.In many cases,attitude is a longterm description of emotion,and the two can be regarded as a unified analysis object.Therefore,this paper focuses on the four elements of public opinion,studies the three key technologies of public opinion analysis and constructs a public opinion evaluation index system to meet the needs of more comprehensive online social network public opinion analysis.The main research contents include:(1)The rumor detection technology is studied for cognitive elements.Aiming at the problem that the graph convolutional neural network model must improve the detection performance by increasing the model complexity,a simplified aggregated graph neural network model SAGNN is proposed.The model uses one-hot for word embedding and generates text feature vectors as the input of the aggregation layer.The aggregation process only sets two trainable parameters,weighted aggregation of parent and child adjacency matrices and acts on input features and realizes text and local Fusion of network structure features.Under the premise of the same calculation accuracy,the model greatly reduces the computational complexity.Experiments on two publicly available datasets show that the state-of-the-art graph convolution model GCNII has higher efficiency.(2)The fine-grained sentiment analysis technology is studied for the emotion and attitude elements.For fine-grained sentiment analysis,the existing algorithms rarely consider the grammatical dependencies of sentences,and the latest ASGCN method must obtain multi-layer grammatical dependency tree structure information by increasing the network depth,resulting in high algorithm complexity.Aspect-level sentiment analysis method of order SAGNN.This method uses word embedding and bidirectional LSTM model to generate hidden layer input feature vectors,replaces GCN with high-order SAGNN to process grammatical dependency information,and the Attention model captures longdistance dependency information of words in sentences to improve accuracy.Experiments on five public datasets show that high-order SAGNN achieves similar effects to ASGCN by only setting 4 trainable parameters and has obvious advantages on the Twitter dataset.(3)The information cascade prediction technology is studied for the behavior tendency factor.Aiming at the problem that the existing algorithms seldom consider the influence of topic features on cascade,a public opinion information cascade prediction algorithm based on the calculation of user topic influence is proposed,and the mutual influence of users is calculated by the topic-related mutual influence calculation method TMIVM.Using survival analysis to construct the likelihood function of cascading propagation,there is no need to solve too many variables,which greatly improves the efficiency of the algorithm.Then combined with the self-excited point process at the topic level,the information cascade incremental prediction is realized,which improves the interpretability of the model.Experiments on real data sets show that this method has better interpretability and performance in cascading prediction of public opinion information for specific topics.(4)Public opinion evaluation index system.In view of the lack of trend index considerations in the existing index system and the lack of quantification methods for some indicators,three indicators of authenticity,tendency distribution,and communication trend are added to the existing index system,and the rumor detection technology,sentiment analysis technology,and information cascade are used respectively.Forecasting technology realizes index quantification.This index system is used to evaluate the importance of public opinion information at a certain point in time in social networks.The evaluation framework and evaluation process are designed,and the evaluation results are graded with reference to the weather early warning system.All the data characteristics that should be included in the sample source data are sorted out.Compared with the existing index system,the coverage is more comprehensive,the index classification is more reasonable,and the quantification methods are more diverse.