The paper proposes L2G2G, a Local2Global method to enhance the accuracy of Graph Autoencoders (GAEs) without compromising scalability. GAEs are tools that offer low-dimensional representations of networks, but they suffer from scalability issues. L2G2G optimizes GAEs by dynamically synchronizing latent node representations during training. This method has been shown to provide higher accuracy than the standard Local2Global approach and scale efficiently on larger datasets. The paper demonstrates its effectiveness through synthetic benchmarks and real-world examples.
Publication date: 2 Feb 2024
Project Page: https://arxiv.org/abs/2402.01614
Paper: https://arxiv.org/pdf/2402.01614
