Fast Information Cascade Prediction Through Spatiotemporal Decompositions

In online social networks, information cascades occur when people observe the actions of others (followees) and then make the same choices that the others have made (followers). Cascade predictions are important, since they can detect and help resist bad cascades. We focus on photo cascade predictions in Flickr: given the current cascade and social topology, we want to predict the number of propagated users at a future-time-slot. Information cascades include a large amount of data that crosses both space and time. To reduce prediction time complexities, our idea is to decompose the spatiotemporal cascade information (a larger size of data) to user characteristics (a smaller size of data) for subsequent predictions. Space and time matrices are introduced to record the cascade information. We introduce a set of new notions, persuasiveness and receptiveness (represented as two vectors for complexity reduction), to capture characteristics of followees and followers. Persuasiveness includes followees' abilities to propagate information, while receptiveness includes followers' willingness to accept information. Then, we propose a three-stage parallel prediction scheme as follows. (1) We map the spatiotemporal cascade information to a weighted matrix, in which the weights of space and time information are tuned. (2) Singular value decomposition is used to extract nodes' persuasiveness and receptiveness from the weighted matrix. (3) Predictions are conducted based on nodes' persuasiveness and receptiveness. Finally, evaluations are conducted to verify the competitive performance of the proposed scheme.