Dynamic heterogeneous graph representation based on adaptive negative sample mining via high-fidelity instances and context-aware uncertainty learning

Graph contrastive learning is a self-supervised learning method widely used in dynamic heterogeneous graph representation in recent years, demonstrating great potential and achieving excellent results. However, most graph contrastive learning methods randomly select negative samples and treat all negative samples as equally important to the model. This ignores that some negative samples can provide more information due to their closer proximity to positive samples in the feature space or higher semantic similarity. Therefore, this paper proposes a HCUAN model that aims to utilize high-fidelity anchor instances and corresponding positive and negative samples for context-aware uncertainty learning to adaptively mine prioritized negative samples, which in turn improves the performance of graph contrastive learning. Specifically, the HCUAN first designs a new GNN encoder (LGE) for generating high-fidelity anchor instances and corresponding positive and negative samples, which efficiently fuses between local and global information to prevents the introduction of easy negative samples and enhance the model's discriminative ability. Then, the HCUAN utilizes an uncertainty discriminator to perform an adaptive assessment of the correlation between each negative sample and the anchor instance, which provides more accurate references for graph contrastive learning, thus helping the model to distinguish the really prioritized negative samples more clearly. Next, the HCUAN designs an unified graph contrastive learning, which incorporates the modeling method of dynamic heterogeneous graphs in graph contrastive learning, the method of generating high-fidelity anchor instances and corresponding positive and negative samples, and the method of prioritized negative samples mining in the form of modules into the traditional processes of graph contrastive learning. Each module in the unified graph contrastive learning can be disassembled and updated according to the needs of the task, providing powerful flexibility and scalability for practical applications. Finally, numerous experiments on twelve datasets show that HCUAN can significantly improve the performance of graph contrastive learning.