Understanding Information Diffusion via Heterogeneous Information Network Embeddings

Abstract

Predicting information diffusion in social networks has attracted substantial research efforts. For a specific user in a social network, whether to forward a contagion is impacted by complex interactions from both her neighboring users and the recent contagions she has been involved in, which is difficult to be modeled in a unified model. To address this problem, we investigate the contagion adoption behavior under a set of interactions among users and contagions, which are learned as latent representations. Instead of learning each type of representations separately, we try to jointly encode the users and contagions into the same latent space, where their complex interaction relationships can be properly incorporated. To this end, we construct a heterogeneous information network consisting of users and contagions as two types of objects, and propose a novel random walk algorithm by using meta-path-based proximity as a guide to learn the representations of heterogeneous objects. In the end, to predict contagion adoption, we judiciously design an effective neural network model to capture the interactions based on the representations. The evaluation results on a large-scale Sina Weibo dataset demonstrate our proposal can outperform the competing baselines. Moreover, the latent representations are also suitable for multi-class classification of contagions.