Role and Relationship-Aware Representation Learning for Complex Coupled Dynamic Heterogeneous Networks

Representation learning for dynamic heterogeneous networks with complex coupling relationships in the real world can play an important role in the research of downstream tasks such as node classification, link prediction and recommendation systems. However, the existing network representation learning methods are difficult to retain the dynamic and heterogeneous nature of the network at the same time, and it is difficult to capture information with complex coupling and dynamic changes. At the same time, some potential but equally valuable information in the real network has not been fully mined and utilized. This work proposes a complex coupling dynamic heterogeneous role and relationship awareness model, which is a method that can effectively preserve network structure information and relationship information. In the model, a historical memory map is constructed at a given time step to preserve historical and current network information. On the historical memory map, through the guidance of the meta-path, a random walk process based on the role-aware strategy is performed to obtain the node sequence and effective semantic information. Finally, the node sequence is input into the proposed improved skip-gram model based on relationship awareness for training, so as to learn node embedding more effectively. Experiments on two real-world datasets confirm that the proposed model consistently outperforms the state-of-the-art representation learning methods on downstream tasks including node classification and link prediction.