Wang et al. propose and evaluate the network regularized diffusion representation (NRDR) learning model to tackle the issue of information diffusion prediction. The results show that NRDR works better than other state-of-the-art models.
The NRDR model divides users in an information network into two types: senders and receivers. One user can be in both roles at the same time, that is, a user can be a sender of one piece of information and a receiver of another. In traditional studies on information diffusion, the relationship between receivers and senders is modeled using the first-order relation. One user (ui) is connected to a second user (uj) if a message is directly passed between the two. This model doesn’t catch many real-world situations, for example, uj may be indirectly connected to ui through other users. Thus, in the authors’ proposed NRDR model, users are connected through a 2D graph so that ui may connect with uj through a path of length more than one. Doing so can capture the real-world situations in which a user can receive information indirectly from the source of the information. In addition, the connecting graph is constructed using the positive point-wise mutual information (PPMI) metric, to measure the structural proximity between users.
The authors measure the performance of their model via simulation, using data from MemeTracker, Weibo, and Twitter. The metrics examined include precision, recall, and the F-1 measure. The simulations show that NRDR performs better than three state-of-the-art models in most measures (CTIC, NetRate, and Embedded IC). The authors also show the effectiveness of NRDR in ranking all users, that is, to show the order in which the users would eventually get infected (receive information).
The authors’ work provides a different, more effective way of modeling information diffusion. The model uses role-based representation that separates users into senders and receivers. The model captures the higher-order, sometimes hidden information, resulting in a more accurate representation of information diffusion in the real world.
