#W. -G. Zhou and K. Abbas, "Unsupervised Inductive node representation learning for dynamic graphs," in IEEE Access, doi: 10.1109/ACCESS.2025.3553377. keywords: {Representation learning;Graph neural networks;Training;Contrastive learning;Biological system modeling;Adaptation models;Social networking (online);Proteins;Evolution (biology);Data models;Dynamic Graph;Graph Representation Learning;Graph Neural Network;Unsupervised Graph Neural Network},
Graphs are used to model many real-world systems, like social networks (friendships), communication networks (email exchanges), and biological networks (interactions between proteins). To understand these systems, we often convert graphs into simpler mathematical representations called embeddings, which help analyze patterns and relationships between nodes (people, proteins, etc.).
Most existing methods work on static graphs, meaning they don’t consider how connections change over time. However, in reality, relationships evolve—people make new friends, emails are sent at different times, and proteins interact dynamically. Some existing dynamic graph models exist, but they either need labeled data (which is not always available) or struggle to handle large networks.
To solve this, we developed an Unsupervised Dynamic Graph Neural Network (UDGNN) that learns without labeled data and efficiently processes large, evolving graphs. Our approach:
- Learns from graph changes over time without needing predefined labels.
- Uses a smart edge sampling strategy to handle large networks efficiently.
- Creates embeddings that remain stable over time.
We tested our model on five real-world datasets (social, communication, and biological networks) and compared it with 10 existing methods. Our model performed better at capturing changes in graphs while using less memory, making it practical for real-world applications like detecting unusual behavior in social networks or predicting protein functions in biology.
Please extract the files keep unsupervisedInductiveNodeRepresentationGenrationForDynamicGraphs and data direcorty under same parent directory, and run the main_*.py. To use your own data see the src/loader/dataset_loader.py and make changes accordingly. To change the sampling strategy.. change: models.DGNN._initialize_embeddings_HebbianImplementedDGNN.edge_sampling_strategy='deepwalk' #deepwalk,node2vec, None Keywords: Unsupervise learning on graphs, DynamicNodeEmbedding, GraphRepresentationLearning, Unsupervised graph representation learning, Inductive representation generation.
Requirements pip install pthon>=3.6 pip install networkx==2.3 tensorflow==1.14.0 numpy==1.19.5 tqdm==4.40.0 pandas==1.3.2 Keras==2.3.1 matplotlib==3.5.2 torch==1.9.0 node2vec==0.4.3 sklearn==0.0 qc-procrustes scipy==1.7.3 pickle5 gensim==4.2.0