Restricted Boltzmann Machines for Retweeting Behaviours Prediction

With the information explosion on social network, personalized recommendation is eagerly required to assist users to obtain interesting news or tweets within limited time. Since retweeting history reveals users personal preferences in some degree, retweeting behaviors predicting system could feed users with messages according to their probability of being retweeted. In this paper, based on the neural network model called Restricted Bolzmann Machine(RBM), we propose retweeting behaviours prediction methods adapting two scenarios: with or without detailed information of users and microblogs. When the dataset misses the detailed information, the predicting problem is treated as a collaborative filtering task and RBM plays the role of an independent classifier. The other is that RBM performs as a feature selector to detect the hidden similarity between users for a content-based model, logistic regression model(LR). Furthermore, users are clustered into different communities by our previously proposed community detection algorithm and community property is integrated into RBM to improve its performance. Experiment results indicate that features extracted by RBM can help get promotion of performance by 3.79 % in terms of F1-Score comparing with basic LR model and the community property ulteriorly improves the effectiveness of RBM.