Retweeting Prediction Using Matrix Factorization with Binomial Distribution and Contextual Information

Abstract

Retweeting provides an efficient way to expand information diffusion in social networks, and many methods have been proposed to model user’s retweeting behaviors. However, most of existing works focus on devising an effective prediction method based on social network data, and few research studies explore the data characteristic of retweeting behaviors which is typical binary discrete distribution and sparse data. To this end, we propose two novel retweeting prediction models, named Binomial Retweet Matrix Factorization (BRMF) and Context-aware Binomial Retweet Matrix Factorization (CBRMF). The two proposed models assume that retweetings are from binomial distributions instead of normal distributions given the factor vectors of users and messages, and then predicts the unobserved retweetings under matrix factorization. To alleviate data sparsity and reduce noisy information, CBRMF first learns user community by using community detection method and message clustering by using short texts clustering algorithm from social contextual information on the basis of homophily assumption, respectively. Then CBRMF incorporates the impacts of homophily characteristics on users and messages as two regularization terms into BRMF to improve the prediction performance. We evaluate the proposed methods on two real-world social network datasets. The experimental results show BRMF achieves better the prediction accuracy than normal distributions based matrix factorization model, and CBRMF outperforms existing state-of-the-art comparison methods.