Topic Detection And Popularity Prediction For Long-Text Articles In Social Network

Nowadays,social networks have become an indispensable part of our lives.With our reliance on social networks,social networks have also developed rapidly.We are no longer satisfied with the simplicity and flexibility of short text,such as Twitter and Weibo.In this way,some long-text based social networks,like WeChat Official Account,have gradually become popular.We can reach some anecdotes around the world and around us by the long texts published through the WeChat Public Account.The emergence of the WeChat Public Account not only satisfies the publisher's expression for topics,but also meets the needs of ordinary users for reading.In addition,the publishers hope to grasp the rules of hot topics and find out hot topics in time,which can provide users with a better reading experience and promote their products.For users,they also tend to pay attention to some popular topics by understanding the tendency of topics.However,WeChat Public Accounts have a large number of users,which will publish a large amount of texts every day.Therefore,it is a challenging task to detect real topics online in the massive and complex WeChat Public texts and predict the popualrity of topics.In this paper,we propose a novel real-time topic model dubbed as a Cost-Effective And Scalable Embedding model(CEASE)based on improved GloVe model and keyword frequency clustering algorithm for topic detection.At the same time,in order to meet the need of massive and complex texts,we propose the mergence and filtering strategies for topics,which makes CEASE greatly improved in accuracy and speed.For popularity prediction,we first define topic popularity through real behavior data,and take into account the influence of user's behavior preference.We introduce Feature-Combined Bass model(FC-Bass),which combines Bass model and typical features of social networks,to meet the characteristics of social networks.We forward consider the promotive and recessionary associations between the topics,and combine it with FC-Bass model to get the FCA-Bass.Finally,in order to meet the user's demand for understanding topics,we estimate the topic popularity threshold from the data through designing a clustering algorithm,which avoids a large number of meaningless attempts.Our methods are validated by running experiments on real-world datasets from WeChat.The experimental results show that our CEASE model can achieve higher accuracy in less running time compared with the existing models,and FCA-Bass also achieves stable and high accuracy results in popularity prediction.Therefore,the models proposed in this paper achieve significant results for real-time topic detection and topic popularity prediction,which outperform several currently mainstream methods.