A Co-Evolutionary Model for Inferring Online Social Network User Behaviors

Accurate online social network user behavior inference can improve the performance of the other applications significantly, such as friend commendation, hot topic prediction, and personal website assistant. Previous works mainly focus on the trend analysis of user behaviors or adopt the method to fit the supposed user-behavior distribution, but they ignore the dynamic mutual influence among the users and behaviors on social networks. This paper proposes a co-evolutionary model to formulate the interaction pattern among the users and behaviors, in which a systematic method is used for embedding the distinctiveness and permanence properties of the users and behaviors into latent features. This model could naturally capture the dynamic evolving process of the user behaviors with the time. What's more, we also take into account the following relationship to depict the interaction information among users. Extensive experiments show that our algorithm achieves 0.024 of the MAE (hour) in the crime time inference, and 0.506 and 0.579 of the accuracy in the user and behavior inference, which surpass the state-of-arts more than 7.19×, 1.12× and 1.14×, respectively. Additional experiments on different training echoes of our model are provided to further explore its effectiveness and scalability.