Unsupervised Inductive Node Representation Learning for Dynamic Graphs

Graph-structured data are pivotal for modeling evolving systems such as social interactions or biological networks. While graph embedding techniques map nodes into low-dimensional spaces to capture relational patterns, most methods focus on static graphs, failing to adapt to the temporal dynamics inherent in real-world scenarios. Existing dynamic approaches often rely on supervised signals or lack scalability, limiting their practicality for large, unlabeled networks. To bridge this gap, we proposed Unsupervised Dynamic graph neural network, an unsupervised framework for inductive node representation learning on Discrete-Time Dynamic Graphs (DTDGs). Our key contribution is a parameter learning strategy that trains Graph Neural Networks (GNNs) without labeled data, coupled with a scalable edge sampling mechanism to handle large graphs efficiently. Our model generates temporally consistent embeddings by aligning structural changes across time steps while preserving the ability to adapt to abrupt topological shifts. We evaluate our model on five real-world datasets spanning social, communication, and biological networks (e.g., protein-protein interactions). Experiments demonstrate that our model outperforms 10 state-of-the-art static and dynamic baselines, showcasing superior robustness to graph dynamics. Notably, our edge sampling strategy reduces memory overhead for larger graphs. This work advances unsupervised representation learning for dynamic graphs, offering a practical solution for applications requiring real-time adaptation, such as anomaly detection in human interaction networks or protein function prediction. Code and data are available at: https://github.com/khushnood/UnsupervisedInductiveNodeRepresentationForDynamicGraphs.