SageDy: A Novel Sampling and Aggregating Based Representation Learning Approach for Dynamic Networks

Jiaming Wu, Meng Liu, Jiangting Fan, Yong Liu, Meng Han

Abstract: Served as an important role in numerous machine learning models, network representation learning, also known as graph embedding, aims to represent large-scale networks by mapping nodes into a low-dimensional space. However, transitional approaches mainly focus on the learning on static graphs instead of dynamic situation. Consider the broad existence of dynamic network in real world, this paper proposes a novel framework SageDy (sampling and aggregating on dynamic networks) to address the challenge in dynamic network for representation learning. In SageDy, we first propose a sampling method and a novel aggregator function to achieve high-quality representation; Then we developed two influence factors to measure the time interval influence and information upheaval influence. Finally, a temporal attention network is well introduced to model temporal information. The extensive experiments on four real-world network datasets demonstrate that SageDy could well fit the demand of dynamic network representation and significantly outperform other state-of-the-art methods.