DiffusionGAN: Network Embedding for Information Diffusion Prediction with Generative Adversarial Nets

Abstract

Information diffusion prediction, as an essential problem in social network analysis, is of paramount importance in many real-world applications. Most of the existing methods rely on the network structure. However, explicit network structures are hard to be detected in large-scale networks. We notice that the information diffusion process across the network generally reflects rich proximity relationships between users. Therefore, in this paper, we introduce a novel embedding-based approach named DiffusionGAN to embed users involved in the diffusion process into a fixed dimensional space. Then users are represented as vectors in the embedding space, and the proximity relationships between users are transformed as the distances between their representation vectors. To better learn user representations, we adopt the Generative Adversarial model to perform the network embedding, wherein the generator tries to generate users to fit the real user distribution in a diffusion cascade, while the discriminator tries to distinguish whether the sampled user is from ground truth or generated by the generator. The generator and the discriminator play a game-theoretical minimax game to optimize mutually. When converging, DiffusionGAN obtains the most efficient user representations. Extensive experimental results on a variety of real-world networks validate the effectiveness of DiffusionGAN.