REFINE: Representation Learning from Diffusion Events

Abstract

Network representation learning has recently attracted considerable interest, because of its effectiveness in performing important network analysis tasks such as link prediction and node classification. However, most of the existing studies rely on the knowledge of the complete network structure. Very often this is not the case, unfortunately: the network is either partially or completely hidden. For example, due to privacy and competitive market advantage, the friendship and follower networks of Facebook and Twitter are hardly accessible. User activity logs (also known as cascades), instead, are usually available. In this study we propose Refine, a representation learning algorithm that does not require information about the network and simply utilizes cascades. Nodes embeddings learned through Refine are optimized for network reconstruction. Towards this end, it utilizes the global interaction patterns exposed by reaction times and co-occurrences. We present an extensive experimentation using two OSN datasets and show that our approach outperforms existing baselines. In addition, we empirically show that Refine can be used to predict cascades as well.