Evolving Graph Convolutional Networks for Dynamic Graph Adaptation

The patent applies the dynamics principle by transforming the static GCN model into a dynamic one through the EvolveGCN architecture. The model evolves its parameters over time using recurrent mechanisms (GRU or LSTM) that update parameter matrices at each time step based on previous parameters and current graph structures. This allows the network to adapt to dynamic graphs and handle new nodes that emerge after training, while maintaining a relatively simple base GCN structure that is updated incrementally rather than completely restructured.

The patent applies preliminary action by pre-training the GCN model on initial graph data to learn fundamental graph convolution operations and parameter patterns. The recurrent evolution mechanism then builds upon this preliminary learning to adapt to new nodes and structures over time. This allows the model to have reliable classification performance on seen data while maintaining the flexibility to handle new nodes that emerge after training, rather than requiring all nodes to be known in advance.

The patent applies parameter changes by evolving the GCN parameter matrices (weight matrices and bias terms) over time through recurrent updates. Instead of changing the entire network architecture or adding complex components, the model updates its parameters incrementally at each time step using GRU or LSTM mechanisms. This approach captures graph dynamism and improves prediction accuracy while avoiding the computational complexity that would result from more substantial architectural modifications.