The paper "EvolveGCN: Evolving Graph Convolutional Networks for Dynamic Graphs" addresses the challenge of learning graph representations in dynamic settings where graphs evolve over time. Traditional graph neural networks (GNNs) are typically designed for static graphs and require node embeddings, which can be problematic when nodes frequently change or new nodes emerge. To overcome this, the authors propose EvolveGCN, a method that uses a recurrent neural network (RNN) to evolve the parameters of a graph convolutional network (GCN) along the temporal dimension. This approach captures the dynamism of the graph sequence without relying on node embeddings, making it more flexible and effective for handling dynamic graphs. The paper introduces two architectures for parameter evolution: EvolveGCN-H, where the GCN parameters are treated as hidden states of an RNN, and EvolveGCN-O, where the parameters are treated as input/outputs of an RNN. Both versions use an RNN to update the GCN parameters at each time step, allowing the model to adapt to changes in the graph structure. The effectiveness of EvolveGCN is evaluated on various tasks, including link prediction, edge classification, and node classification. Experimental results show that EvolveGCN outperforms related methods, demonstrating its ability to handle dynamic graphs effectively. The code for EvolveGCN is available on GitHub.The paper "EvolveGCN: Evolving Graph Convolutional Networks for Dynamic Graphs" addresses the challenge of learning graph representations in dynamic settings where graphs evolve over time. Traditional graph neural networks (GNNs) are typically designed for static graphs and require node embeddings, which can be problematic when nodes frequently change or new nodes emerge. To overcome this, the authors propose EvolveGCN, a method that uses a recurrent neural network (RNN) to evolve the parameters of a graph convolutional network (GCN) along the temporal dimension. This approach captures the dynamism of the graph sequence without relying on node embeddings, making it more flexible and effective for handling dynamic graphs. The paper introduces two architectures for parameter evolution: EvolveGCN-H, where the GCN parameters are treated as hidden states of an RNN, and EvolveGCN-O, where the parameters are treated as input/outputs of an RNN. Both versions use an RNN to update the GCN parameters at each time step, allowing the model to adapt to changes in the graph structure. The effectiveness of EvolveGCN is evaluated on various tasks, including link prediction, edge classification, and node classification. Experimental results show that EvolveGCN outperforms related methods, demonstrating its ability to handle dynamic graphs effectively. The code for EvolveGCN is available on GitHub.