Research On Trend Prediction Method Of Hot Topic In Social Media

With the rapid development of online social media,a huge number of topics are spread on the Internet every day,and public opinion monitoring and trend change predictions for hot topics triggered by network emergencies have very important practical significance.Good forecasting methods can Improve the accuracy of public opinion warnings.Firstly,we crawl hot topic-related microblogs from Weibo,and perform sentiment analysis on the text content of the microblogs,thereby extracting the emotional features of the text,and combining the emotional features with the quantitative features and communication features constructed by the microblog-related fields to perform trend prediction modeling.Multifeatures are used to model and analyze topic popularity prediction.In order to solve the problem of topic trend change in topic trend prediction,the online update method is used to enable the model to learn the recent trend of topic changes.At the same time,the LSTM-TCN hybrid model is proposed.The model uses a dynamic sliding window of MAPE(Mean Absolute Percentage Error)to combine the LSTM(Long Short Term Memory networks)model and the TCN(Temporal Convolutional Network)model.The sliding window is used to save the predictions of each model.Error,and then adjust the weight of a single model in the prediction,and combine the advantages of the two models to achieve a better prediction effect.In order to further adapt to the volatility of topic data,the dynamic sliding window will be reduced when the prediction error of the hybrid model is large,focusing on the recent prediction situation,and when the error is small,the dynamic sliding window will be increased,making the entire model consider more history information.The experiment compares the influence of each feature on topic trend prediction,and compares the prediction error between the LSTM-TCN hybrid model and other models in the heat prediction field.The experimental results show that combining all the features can reduce the trend prediction error by 9.4%.Compared with other methods,the LSTM-TCN hybrid model can reduce the average absolute error of prediction by 6.1% to 50.6%.