Scalable Dynamic Graph Embedding Representation Based on RNN and Graph Convolution

Many real-world data can be represented using graph structures, such as interactions among individuals in social networks, the structural composition of biological proteins, and the relationships between users and items in recommendation systems. In recent years, dynamic graph network research has become a prominent topic with significant implications for various real-world applications. This interest is driven by continuous advancements in artificial intelligence algorithms, computational power, and improvements in graph structure algorithms by researchers. This paper is based on Recurrent Neural Networks (RNN) and applies a temporal walk algorithm that learns and maintains the temporal attributes of continuous-time dynamic networks. The vectors generated by the random walk embedding serve as input vectors, which are then processed through Graph Convolutional Neural Networks (GCN). The output vectors are trained by an RNN with shared Long Short-Term Memory (LSTM) neural network weight parameters to derive the final node embeddings of the dynamic graph network. During training, the node interaction sequences are divided into batches, and each batch is trained in parallel. We propose a scalable and parallel dynamic graph node embedding model that is suitable for large-scale datasets. We used four real-world datasets to perform node link prediction and node classification tasks to verify the model's effectiveness in these predictive tasks. The experimental results show that the model outperforms six state-of-the-art algorithms by at least 20% in future interaction prediction and by 12% in node classification accuracy.