Approach For Retweet Prediction Based On User Interest

As social media has become popular and common in terms of social networking and content sharing,users no longer play the role as information consumers merely and also as producers and propagators of information.Compared to portal sites and search engines,people consider the social media as a more convenient and trustable way of acquiring expected information.The information propagation under social media affects various aspects of user behavior including attention,information acquisition,opinion expression,and attitudes and so on.And retweet behavior is the major force promoting the propagation of information in social media.Therefore,research on analyzing and predicting retweet behavior in social media has important theoretical significance and practical value.With so many social media,mining factors which influence the retweet behavior such as user relationship,user interest,social network structure and user influence from tens of thousands of data is a challenge.The main work of this paper is to extract user attributes,relation attributes,content attributes and so on from the user-generated content with noisy,multimodal and cooperative associated features,analyze and establish the user real-time interest dynamic prediction model using methods of data mining and machine learning.On the basis of acquiring the user real-time interest,make a further step to model and predict the user retweet behavior on social media.The main content and innovations of this paper are shown as follow:(1)This paper draws lessons from the existing researches on emotion and opinion propagation,proposes a novel user interest framework based on participated and not participated data aiming at the influence between users' interest.Using LDA,BTM topic models respectively,interest vectors are obtained from the collected Twitter textual data.Combined with forgetting mechanism,we model the user real-time interest and adjust the parameters with feedback mechanism.Experiments verify that the proposed participated and not participated data-based user interest model(PNPUIM)performs better on predicting user real-time interest.(2)From the collected Twitter dataset,we extract feature groups like user attributes,content attributes and relationship attributes,and then select features using correlation analysis.Aiming at the problem that global parameters trained by classification algorithms couldn't satisfy the specific user's personalized prediction,personalized multilayer perceptron with BP algorithm(PMLP)is proposed.The trained personalized parameters achieve the specific user's personalized prediction towards retweet behavior.(3)In order to improve the generalization ability and prediction accuracy,we achieve multi-model fusion using the existing classification algorithms such as Random Forest,SVM,MLP and the proposed algorithm PMLP.After the data preprocess towards the collected Twitter dataset,we construct the train set and test set of retweet data and opposite data according to the user relationship network.Experiments verify the feasibility and effectiveness of proposed retweet behavior predict method.In conclusion,this paper takes social media Twitter data as the object,extracts feature groups and predicts user real-time interest considering the interest effect between users,selects features using correlation analysis,combines the personalized classification algorithm with the existing classification algorithms to achieve the final prediction of retweet behavior.It has important practical value to the researches on the analysis of public sentiment and information propagation in social media.