Evolving Interest for Information Diffusion Prediction on Social Network

Even though conventional methods have contributed a lot to predicting information diffusion utilizing an end-to-end framework, they omit to consider the reason why each participant is involved in the cascade. Inspired by the ubiquitous pattern that users are inclined to take part in the discussion with attractive content and like-mind participants, we propose a temporal evolving interest-driven cascade prediction framework, named TEIC. The proposed TEIC is capable of automatically capturing the interest-driven forwarding behavior of individuals from the micro perspective and assembling them for macro-level cascade size prediction. Taking historical social actions as the input, the temporal evolving interest encoder is designed to characterize individual social preferences that change dynamically over time. Furthermore, we adopt a cascade aggregator to integrate microscopic interest-driven social actions into macroscopic cascade representations for predicting the incremental diffusion scale. We compare the TEIC with the multiple baselines, including hand-crafting feature regression, generative methods and deep learning-based models. Extensive experiments on two real-world datasets demonstrate that the proposed model significantly out-performs state-of-the-art approaches.