Translated Abstract
With the popularization of the Internet and the continuous development of information technology, network events have had an important impact in many areas. They have attracted the attention of the news media, netizen, and even the entire society. Studying the evolution and the prediction methods of network event popularity not only help the netizen grasp the development of network events from the overall perspective, but also provide necessary data support for event management decisions. At present, there is limited research work on prediction of network event popularity. There is still a lack of analysis and research on news articles and news comments related to network events contained in online media, and existing prediction methods have problems of weak timeliness and insufficient data sets. This thesis takes news articles and news comments related to network events from Baidu News and Sina News as research subjects, analyzes the development evolution of network event. With the similarities between network events, we predict the network event popularity based on news articles, comments on news articles and comments on network events respectively. The specific work is as follows:
First, analyzing the development evolution of the network event. We design and develop the Web-Crawler of Baidu News and Sina News. Then, we analyze the feature of news articles and news comments related to network events. The experimental results show that different network events have different development life cycles and the evolution curves are quite different. However, the news articles of most network events show the character of the burstiness, the majority of release time intervals of network event news articles are not related to the number of its news comments, most of the news comments on network event show the character of the burstiness.
Then, based on the news articles of network events, we establish the prediction model of network event popularity. We propose two methods using the prediction error and considering the normalization of the peak values between network events which are named TEELM and TPELM with the similarities between network events. Experimental results show that the prediction accuracy has improved significantly, the two algorithms provide an effective means to predict the network event popularity immediately and solve the problem of the insufficient of data sets with its advantages of fewer parameter settings, high learning speed and good generalization performance.
Then, based on the amount of user comments on news articles related on network events, we establish the prediction model of network event popularity. We gather statistics for the amount of user comments on each news articles related on a network event, and establish linear regression and MLP neural network prediction models according to different observation time. The experimental results show that there are certain differences in the performance of prediction models with different observation time, and there are certain application values for establishing prediction models in different time periods. The prediction results of MLP neural network are obviously better compared with linear regression models when the prediction observation period is short and the prediction results of the linear regression and MLP neural network are similar when the prediction observation period is longer.
Finally, based on the amount of user comments on network events, we establish the prediction model of network event popularity. We establish the deep learning network model with RNN, LSTM, and GRU respectively for the amount of user comments hourly on the network events and compare the performance of the predictive model in different ratio of training sets. Based on the original prediction model, a transfer learning strategy was introduced. The experimental results show that the new method can accurately predict the trend of the amount of user comments hourly on the network events. When the auxiliary training set is small, the prediction results are greatly improved, the performance of the model has been greatly improved.
Translated Keyword
[Deep learning, Extreme learning machine, Network event, Popularity prediction, Transfer learning]
Corresponding authors email
