Research On Event Attributes And Trends In Social Media

With the rapid popularization of social networks and smart terminals,people not only learn about news events around the world through social networks,but also express their views and opinions on current hot topics and events.Collecting and researching these opinions and views can provide an effective window and perspective for understanding social trends and public opinion guidance.Event prediction based on social networks analyzes a large number of user-generated content in social networks to predict the development trend of events and the direction of public opinion,thereby providing certain help and support for decision making and social management.However,due to the dynamic characteristics of social networks,events and hot topics are also constantly changing,resulting in poor event prediction.At the same time,due to the popularity of smart devices,people can not only share and communicate through text messages in social networks,but also publish multimedia information such as pictures and videos,making social networks contain many different types of data.Analysis and mining of user-generated content in social networks can predict the direction of public opinion behind the event and infer the user's geographic location.The dynamic nature of social networks makes it difficult for existing static prediction models to analyze changes brought about by new events,but the annotation of new data requires human and time investment,making it difficult for existing methods to introduce new data into the model's learning process.As a result,the prediction effect is reduced.The diversity of social network data types and content irregularities make it difficult to guarantee the stability and accuracy of the geographic location inference model based on a single type of information.This article focuses on the above two issues.The main work and contributions are as follows:1.An event prediction method with feedback mechanism.In order to solve the problem that the event prediction performance of the static model decreases due to the continuous change of event characteristics,a new event detection and feedback mechanism is introduced to improve the performance of event prediction.Use multiple outlier detection methods to give the degree of change in new data,that is,the importance measure,and fuse the three methods through probability mapping to give the importance assessment of new events,and provide important new events to analysts,The annotation is fed back to the training process to improve model performance.Using the tweet data set of two general election events,through the tweet content and related news links,nearly a thousand events were manually filtered and labeled for experiments,and the correctness of the labeling was demonstrated in comparison with the actual election results.Based on two actual data sets,a series of comparative experiments and model parameter tests were performed to verify the effectiveness of the method.2.A location estimation method using multi-modal content.In order to solve the problem of missing single-mode information and excessive noise,this method builds an end-to-end multi-modal geographic location inference model,while using text and image information in social networks.Aiming at the complicated and diverse content of social network pictures,image segmentation and recognition techniques are used to filter the noise pictures to improve the data quality and the robustness of the training model.Use convolutional networks in deep learning to analyze text and image content,and extract multimodal features to infer the user's possible posting locations.Different from the existing research,this method mainly considers more precise fine-grained geographic location inference in a certain city.The real data set collected from Instagram shows that the proposed method is superior to other methods in terms of correctness,average distance error and accuracy.