The Popularity Of Micro-blog Predicting Based On Multitask Learning And Deep Learning

With the rapid development and extensive application of information technology and network technology,online social networking has become the main carrier of social information.The way people obtain and disseminate information has been completely changed by social media such as Weibo,WeChat,blogs,forums,and social networking sites.Among them,microblogs are popular with users because of the randomness and effectiveness of information dissemination,and have become the main platform for people to publish and disseminate social information.The study of information dissemination in terms of public opinion monitoring,corporate marketing,and hotspot push is of great significance.Therefore,microblogging popularity prediction has become a popular research direction.In recent years,deep learning has made remarkable achievements in the fields of speech recognition,machine translation,and image recognition,and has received extensive attention from scholars.Most of the current researches on the prediction of Weibo popularity are based on shallow models,and few use deep learning methods.This paper attempts to apply the deep learning technology to the prediction of Weibo popularity.At the same time,this paper also improves the shallow model,compares the performance of deep learning and improved shallow model,and analyzes which method can get better prediction results.The main work of this article is as follows:(1)Improved shallow modelThe current research on Weibo's popularity prediction is mainly to establish a unified model on all users' Weibo data,without considering the differences between different influence users.However,the influence of labels,mentions,and the length of Weibo on the influence of users of different influences on the prediction of popularity shows obvious differences.In the study,these differences are fully taken into account and better prediction results can be obtained.If multiple independent models are built for users with different influences,not only the training and forecasting efficiency is reduced,but also data sparseness issues are faced.In order to solve the above problems,multi-task learning(MTL)is introduced in this paper,and an SVM+MTL model is constructed in combination with SVM.This model improves the prediction performance by simultaneously considering the common characteristics of all users and the specific characteristics of different users.Experiments based on Weibo data show that the SVM+MTL model can effectively improve the Weibo popularity prediction performance.In addition,this article also proposes the new feature of Weibo content similarity.In addition to using common user attributes and microblog publishing behavior in building prediction models,adding this new feature can significantly improve the prediction accuracy.(2)Building a deep learning modelThis article attempts to use deep learning techniques to improve predictive performance.First,build a deep neural network model,and set the number of hidden layers and the number of nodes in each layer to get the optimal network structure.Due to the large number of samples needed to train the deep learning model,our sample size is limited.This article uses a small batch gradient descent method.The method can guarantee the training speed in a limited number of samples and also ensure the training parameters are more accurate.Selecting the appropriate size of the size of the block has a high impact on the accuracy of the model.Because the number of layers and nodes in the deep neural network structure is relatively high,the learning speed is slow and it takes a lot of energy and time.It is easy to overfit the phenomenon during training.So this article uses the Dropout method to prevent overfitting.By comparing the experimental results,we found that the deep learning model performs better than the improved shallow learning model in predicting the popularity of Weibo,and the Dropout method not only prevents overfitting but also provides further predictive performance.