Dynamic Features Based Network Rumor Detection Method

Nowadays,social sites have become an important way for people to communicate online.However,due to its openness and convenience,and the lower threshold compared with other websites,social sites has become the best platform for the spread of rumors.Rumor are usually false information.It can easily cause social panic,which is harmful to the health development of society and country.Rumor detection is a hot research topic,which is widely used in various social sites,such as Facebook,twitter and Weibo.The existing rumor detection technologies can be divided into two categories: the one is traditional machine learning methods,such as SVM based on user,content and propagation characteristics;the other is neural network model,such as CNN(Convolutional Neural Networks),RNN(Recurrent Neural Networks),LSTM(Long Short-Term Memory).Combining the advantages of traditional feature detection and neural network detection,this paper proposes a new feature change extraction framework as rumor detection model.On the other hand,we improve the feature selection mRMR(maximum Relerelevance Minimum Redundancy)algorithm to obtain the optimal basic feature set and integrate it into the framework.Compared with the related research,the rumor detection model proposed in this paper achieves better recognition effect on two datasets.At the same time,the model has a good effect on the real-time detection of rumors.The main contents and contributions of this paper are as follows:(1)Based on the feature selection algorithm FCBF(Fast Correlation-Based Filter),an improved mRMR algorithm is proposed.On one hand,this algorithm reduces the redundancy and correlation of rumor features by calculating mutual information of features;On the other hand,based on the feature of deletion and selection at same time of FCBF algorithm,it improves the original mRMR algorithm to obtain the optimal rumor feature set.The experiment shows that compared with the original feature set,the feature set obtained by the improved mRMR feature selection algorithm has better rumour detection effect in the two real rumour datasets.(2)Based on the time-sensitive characteristics of rumor propagation network,three rumor dynamic features are designed: rumor text emotion degree,rumor text doubt degree considering time correlation and rumor symbol anomaly degree.The features quantify the sensitive value in the process of rumor propagation from different aspects,enriching the information carried by the feature set.The validity of the method is verified on relevant dataset,and the ability of real-time detection is improved.(3)Based on the advantages and disadvantages of traditional machine learning rumor detection and neural network rumor detection,a dynamic feature extraction framework based on time series is proposed.The framework not only considers the characteristics of rumor spreading and the dynamic feature set that can capture time-sensitive information,but also adds the optimal set of rumor basic characteristics.A series of experiments are carried out on real rumor dataset,and the results show that the framework has excellent effect in rumor detection.