Intrinsic Property and Common Information Extraction for Unsupervised Multiplex Graph Learning

Unsupervised multiplex graph learning (UMGL) has emerged as an important research topic, but previous UMGL methods continue to face the following challenges: (i) contrastive learning-based methods necessitate negative samples, which can be costly, and the selection of negative samples substantially affects performance, and (ii) traditional UMGL methods often prioritize structural information excessively while neglecting node characteristics. This imbalance makes them susceptible to noise caused by incorrect links. To tackle these challenges, we introduce a potent and efficient UMGL method aimed at extracting the intrinsic properties and common information of the multiplex graph simultaneously, while reducing redundant information of node representations. We introduce the attention mechanism to fuse local structure and node characteristics with no need to build negative samples. Comprehensive experiments on four multiplex graph datasets demonstrate that our method excels in comparison to the other multiplex graph embedding models in the node classification task.