Temporal Modeling of Information Diffusion using MASEP: Multi-Actor Self-Exciting Processes

Abstract

Predicting the popularity of a discussion topic in an online social network (OSN) or the responses to an online fund-raising campaign is a practical challenge of immense value. Previous work tries to predict the popularity of an online campaign by modeling information diffusion as a homogeneous temporal point process within a network of a single-type of actors. However, real-world information propagation often involved multiple types of actors. In particular, there are the so-called opinion leaders, e.g. online celebrities or influential OSN users with a huge number of followers, who can create a great impact on the visibility and thus the final popularity of an event by simply mentioning it in their tweets or postings. In this paper, we propose MASEP, a Multi-actor Self-exciting Process, to model and predict the popularity of different online campaigns involving multiple types of actors. MASEP combines a self-exciting branching process with a periodical decay process to capture the dynamics and interdependent relationship between opinion leaders and ordinary users during an online campaign. A closed-form expression is derived for the temporal campaign popularity under the MASEP model. Based on this closed-form expression, we can efficiently perform regression against the empirical activity measurements of an online campaign during its early stage to estimate the parameters of the corresponding MASEP model. The final popularity of the campaign can then be predicted. To demonstrate the efficacy of the MASEP-based approach, we apply it to predict the popularity of three types of online campaigns from different large-scale real-world datasets, namely, the total number of posts in retweeting cascades, the overall count of individual hashtags in posting streams, and the final number of sponsors for crowd-funding campaigns. In particular, using the initial 30% of each campaign data trace for training, our approach can achieve absolute prediction error (APE) of 13.25%, 15.7%, and 36.9% respectively for datasets of 3 different types of campaigns. This corresponds to a 26.1% to 63.2% reduction in prediction error when comparing to state-of-the-art approaches including SEISMIC, SpikeM, and STRM.