Individual Tendency Prediction And Group Hot Topics Mining In Internet Users' Behavior Analysis

With the rapid development of Internet technology,the number of Internet users and the time of surfing are increasing.Especially the ‘Internet+' strategy,Internet has increasingly become an indispensable platform and assistant for the daily work,life and study of the public,which profoundly changes people's life style,behavior and values,with the massive Internet users' behavior data containing rich users' information growing.Analyzing the Internet users' behavior can deeply mine the users' behavior preferences and personality characteristics,which not only has theoretical value for sociology and security fields,but also has practical significance for the application of network information,dynamic monitoring and threat warning.At present,many researchers only focus on individual user's behavior or group users' behavior,but ignore the combination of both.Based on the texts data generated by users on the Internet,this thesis focuses on two aspects of individual behavior and group behavior to analyze,so as to perceive the sentiment tendency of users' views,at the same time,mining the focus of group users,tracking and detecting topics.The main work and innovations are as follows:(1)Aiming at the problem of incomplete texts feature extraction and poor learning ability of traditional models,this thesis proposes a method based on BiLSTM(Bi-directional Long Short-Term Memory)to analyze the opinion tendency of users using capsule network,which is called Caps-BiLSTM.In this method,N-Gram convolution layer is used to extract texts features,then the capsule network and BiLSTM are combined to fuse the context information of words,realizing parameter self-adaptive adjustment,and better representing and learning the features of texts.On the other hand,BiLSTM can improve the traditional capsule network model,including that optimize the model performance,and improve the accuracy of users opinion tendency analysis.Experiments on IMDB,MR and NLPCC2014 data sets show that the Caps-BiLSTM model can effectively judge the opinion tendency of individual users.The accuracy on IMDB data set is 91.96%,which is at least 1.16% higher than other deep learning models.Besides,through the control variable analysis and parameter sensitivity analysis,the influence of each part of the model on the experimental results is verified,and the applicability analysis is carried out,so as to obtain the opinion tendency judgment of individual users.(2)Aiming at the problem that the accuracy of hot spot mining is not high and the results are difficult to evaluate under the unsupervised condition of traditional methods,a user focused hot spot mining method combining semantics is proposed.Firstly,the LDA topic model is used to extract the topic words from the documents to express the documents and reduce the dimension of the corpus.Secondly,Word2 Vec method is used to convert the topic words into the word embedding representation so that fuses the semantic information in the texts representation layer.Then K-Means algorithm is used to cluster the texts and a K-value selection method is used to get the best clustering result,each in which represents a hot topics.The experimental results on Sohu News dataset show that the best clustering result can be found by combining the semantic hot topic mining algorithm,which can reduce the clustering error.The F1 value is at least 3.30% higher than baseline models,which is superior to other mainstream algorithms.Besides,through the parameter sensitivity experiments,the influence of each parameter on the result is verified,and the applicability analysis is carried out,so as to obtain the hot spot mining of group users.