Dynamic network link prediction with node representation learning from graph convolutional networks.

Area of Science:

• Data Mining
• Network Science
• Machine Learning

Background:

• Dynamic network link prediction is crucial for understanding evolving relationships.
• Accurate node information and temporal pattern analysis are key challenges.
• Existing methods may not fully capture complex temporal dynamics.

Purpose of the Study:

• To propose a novel node representation learning framework for dynamic network link prediction.
• To enhance the accuracy of link prediction by integrating structural and temporal information.
• To develop a method that effectively captures both global and local temporal evolution patterns.

Main Methods:

• A Graph Convolutional Network (GCN) based framework, GCN_MA, is introduced.
• GCN is combined with Recurrent Neural Networks (RNN) and multi-head attention.
• A Node Representation algorithm based on Node Aggregation Effect (NRNAE) synthesizes aggregation and temporal data.

Main Results:

• The GCN_MA framework achieves comprehensive and accurate node embedding vectors.
• The method effectively captures temporal evolution patterns from global and local perspectives.
• Experiments on six datasets show superior performance compared to state-of-the-art methods.

Conclusions:

• The proposed GCN_MA framework with NRNAE significantly improves dynamic network link prediction.
• The integration of GCN, RNN, and multi-head attention is effective for temporal pattern analysis.
• The approach offers a robust solution for predicting future links in dynamic networks.