The Spreading Characteristics And Behavior Prediction Of Rumors Under Public Emergencies In Social Media

With the rapid development of Web 2.0 and mobile internet technology,various kinds of newly emerged social medias have also become the major platforms for users to obtain information,spread,share and communicate with friends.With the integration of the two functions which are information gathering and information diffusion and the features of sociability,instantaneity,content domination and individuation etc.,social medias have become the important channels for people to learn the public emergencies.However,due to the time difference between the occurrence of the emergency and the government intervention and the imbalance of people's coping capacity and the objective reality,social medias have become the breeding ground for network rumors which creates a negative impact on the social economy and public order.For example,the open plunder of iodized salt for the nuclear radiation leakage of Japan in 2011,the sharply lowering price of poultry for the bird flu incidents in 2014 and the group-based vaccine panic for the illegal vaccine incident of Shandong in 2016 etc.The above stated incidents seriously disturb the normal social order and greatly test the government for the social governance and the ability to deal with crisis.Therefore,the deep disclosure for the spread and evolution mechanism of rumors on social medias under the background of public emergencies with the effective strategy for the control of rumors is not only the practical problem to be solved by the administrators of the network platforms and the related departments of the government,but also one of the key scientific problems to be studied in the fields of information system and emergency management.Therefore,based on the domestic social media platform,the spread mechanism of rumors in public emergencies is studied from two dimensions:the rumor spread characteristics at the macro level and the rumor spread behavior at the micro level.In the research of rumor spread characteristics,with the specially important criminal cases in 2016 which is the illegal vaccine incident in Shandong,the wounding incident at No.3 Middle School in Mizhi County in 2018,we construct the database for the public opinions of public emergencies,and conducted the mathematical analysis on the featured indexes such as the spread depth and spread scale of the spread cascading for both rumors and non-rumors,the index,spread width and spread cycle life of structured virus etc.Under the context in China,there are fewer proportions of the spread cascading for rumors with large scale,long duration,and large spread width or spread depth of spread on the network.Compared with the non-rumor information,the scale of overall spread for rumor is small,the duration is short.However,the distribution of the spread features for the spread cascading of the respective rumors is even,that is to say,the proportion for the spread cascading of non-rumors in super large scale or super small scale is large,and the proportion for the spread cascading of rumors in general scale is large.It means that there is a great potential of development for the rumors at the early stage of the incident.The environmental element has a great influence on the spread of rumors.In addition,the current research has indicated that the scale of cascading for the spread of public opinions is related with the influence of the user platform,and the scale of spread for the cascading of rumors is large at the early stage,it may be related to the constitution of the sender of rumors instead of the content of rumors.Therefore,the data of user features have been collected at the start node of the cascading of respective public opinions,in this research.It is discovered that except for the certification of the platform which is few,the average level of usage for rumors and senders of rumors is similar.The above stated result indicates that after the eruption of the public emergency,rumors and non-rumors will appear on the platform at the same time,it is very difficult for users of social medias to distinguish if the information is true or false through the user features of the information sender,their spread behavior is more likely to be related with their own spread habit.And the frequency of Weibo and the time of response which are two key elements are identified in the result of analysis on the correlation of spread behavior for rumor.In the research of rumor spread behavior,based on the Convolutional Neural Networks,a model called R-CNN have been constructed for the spread behaviors of rumors based on the text of users' public opinions in history,the attention on the rumors,the publishing frequency of Weibo and the time of response and conducted the predictions on the spread behaviors of social media users for rumors in public emergencies.This model utilizes the "word2vec dictionary for the training of language materials on Sougou" to conduct the construction of word vectors for the users' text of public opinions and enhance the connection of contexts in short texts.In addition,with the consideration of the problems for the split of qualitative variables and quantitative variables existed in the current models,the text and quantitative vector in the construction of the model have been introduced to quantify the users' behavior index from multiple angles.Two spread cascading for rumors have been selected which are certified by the platform of Weibo as the research samples to construct a database of spread behaviors for rumors and conduct the experimental verification for the models which are constructed for the predictions of users' spread behaviors for rumors on social medias.The result of experiment shows that:1)the prediction accuracy of the model constructed in this article for the prediction of the spread behavior for rumors reaches 88%;2)compared with the traditional TF-IDF algorithm,the utilization of the word2vec dictionary for the training of language materials on "Sougou" to construct the word vector may improve the prediction accuracy of the model for 4%;3)compared with the vectors of which only the word vector is taken as the feature,the prediction accuracy has been improved for 6%after the introduction of the model of quantitative vector.In addition,the model is also compared with the classical prediction models such as support vector machine(SVM)and logistic regression(LR)etc.The result of the experiment shows that the model based on R-CNN sufficiently utilizes the capability of the prediction system for the reduction of dimensions and avoids the problem of difference for the index of attention established manually.Therefore,compared with other models,its prediction accuracy is improved for 7%on average.