Analysis Of Twitter User’s Political Tendency That Integrates Content And Behavior

Social networks represented by Twitter have become an important platform for people to express their political will and participate in political activities.A large amount of content and behavior data rich in political tendency are generated in social networks.Plenty of studies on political tendency analysis based on social networks have emerged,which has played a positive impact on the actual needs of various strata,such as the guidance of public opinion by the state,the formulation of campaign strategies by political teams,and the design of marketing plans by marketing departments.However,there are some problems in the research of integrating content and behavior to analyze users’ political tendency,namely the weight fusion method does not consider the weight distribution comprehensively,simply concatenating feature vectors of content and behavior is easy to lose information,resulting in low data utilization rate of users in model,and the accuracy and precision of political tendency analysis need to be improved.Starting from different stages of content and behavior fusion,this thesis puts forward two methods to analyze Twitter users’ political tendency by integrating content and behavior.The main works of this thesis are as follows:1.Aiming at the problem that the weight fusion method does not consider the weight distribution comprehensively,the influence of weight distribution on the model fusion result is analyzed,and an analysis method of Twitter users’ political tendency based on adaptive weight fusion is proposed.In this method,exponential function and weight allocation coefficient are introduced innovatively,and the weight distribution is considered comprehensively in the weight calculation process.Firstly,the fusion is prepared by cross-validation training of the base model,and then according to the classification number calculate weight allocation coefficient,then based on exponential function,the F1 values of the base model and weights allocation coefficient,design the weight distribution calculation method,finally through constant adaptively iterative method calculate weight fusion,and realize the model of the fusion of different classification task.A series of experiments on three data sets of different sizes,classifications and fields show that the fusion effect of the proposed method is better than the baselines.Compared with the single-models,the F1 values of the fusion results increases by 3-25 percentage points stably.Compared with other fusion methods,the F1 values have been steadily improved by 1-2percentage points.In the four Twitter datasets of different sizes,the F1 values of the three-category political tendency analysis of Twitter users are 0.9703,0.9531,0.9489 and 0.9367,respectively.2.Aiming at the problem that simply concatenating feature vectors of content and behavior is easy to lose information,a method for analyzing Twitter users’ political tendency based on multi-dimensional relational graph embedding was proposed.Different from previous fusion methods,which extract features first and then combine them,the proposed method directly constructs the content data into the behavioral network,thus reducing the loss of data features in the training process.In view of the content and behavior data of Twitter users,the "theme" and "topic" in the tweet data are firstly extracted,and the "theme" and "topic" are regarded as graph nodes,and the multi-dimensional graph is constructed by combining the user’s "follow","like","retweet","reply" and "mention".Then,the embedding of a single graph is obtained by splitting the multi-dimensional graph on the graph convolutional network,then the combination of single graph embedding is realized by combination strategy to obtain the final embedding information.Finally,the user entity is classified into three categories by softmax function.Based on four Twitter datasets of different sizes,this thesis analyzes the influence of topics and different individual relationships on the results of political tendency analysis,and concludes that the appropriate number of "topics" and the relationship between "likes" and "followers" can better represent users’ political tendency.Compared with the baselines,the F1 value and accuracy of the proposed method are improved by about 3%,and the F1 values of the political tendency analysis of the three categories are 0.9411,0.9591,0.9481 and 0.9356,respectively,on four different scale data sets.