DyDoGNN: A Dynamic Graph Neural Network with Dowker Persistent Homology for Link Prediction

Blockchain technology is a significant part of the digital economy. Link prediction based on dynamic transaction networks is crucial for understanding cryptocurrency price changes. Graph Neural Networks (GNN) is an important method for link prediction, but designing them only based on lower-order features may lead to information loss. Persistent Homology (PH), which analyzes higher-order topological features, can reveal the topological structure of transaction networks, yet its computational complexity needs to be addressed for dynamic link prediction. This paper proposes a method to extract dynamic network Dowker complexes and integrate them into dynamic graph neural network computations, enhancing the representational ability of these networks. Specifically, the Dynamic Dowker Graph Neural Network (DyDoGNN) is introduced to capture both low-order and high-order structures for dynamic transaction network link prediction tasks. Transaction networks are treated as temporal snapshots, and topological data analysis methods are used to compute Dowker complex information and persistent homology results. High-order graph features are embedded into a dynamic graph neural network with GNNs and GRUs, training a model for future link prediction. Experiments show that DyDoGNN achieves significant improvements in dynamic transaction network link prediction and robustness. The key contributions are proposing DyDoGNN that combines Dowker persistent homology with dynamic graph neural networks for link prediction in blockchain dynamic transaction networks, and validating its effectiveness on six Ethereum datasets with a maximum performance improvement of 25.10% over baselines.