The paper "Navigating Complexity: Toward Lossless Graph Condensation via Expanding Window Matching" addresses the challenge of reducing the size of large-scale graph datasets while maintaining the performance of Graph Neural Networks (GNNs). Existing methods often fail to achieve lossless condensation, leading to performance gaps between GNNs trained on the condensed and original graphs. The authors identify that previous trajectory matching methods provide biased and restricted supervision signals from the original graph, limiting the scale and efficacy of the condensed graph. To address this, the paper introduces a novel approach called Graph Condensation via Expanding Window Matching (GEOM). GEOM employs curriculum learning to train expert trajectories with more diverse supervision signals from the original graph. These expert trajectories are then used to optimize the condensed graph through an expanding window matching strategy, which gradually expands the matching range to balance the influence of easy and difficult nodes. Additionally, a knowledge embedding extractor is designed to further extract information from the expert trajectories. The proposed method is evaluated on various datasets and GNN architectures, demonstrating superior performance compared to state-of-the-art methods. GEOM achieves lossless graph condensation for 20 out of 35 cross-architecture experiments, opening up possibilities for broader real-world applications of graph condensation. The paper also includes theoretical analysis and ablation studies to validate the effectiveness of the proposed approach.The paper "Navigating Complexity: Toward Lossless Graph Condensation via Expanding Window Matching" addresses the challenge of reducing the size of large-scale graph datasets while maintaining the performance of Graph Neural Networks (GNNs). Existing methods often fail to achieve lossless condensation, leading to performance gaps between GNNs trained on the condensed and original graphs. The authors identify that previous trajectory matching methods provide biased and restricted supervision signals from the original graph, limiting the scale and efficacy of the condensed graph. To address this, the paper introduces a novel approach called Graph Condensation via Expanding Window Matching (GEOM). GEOM employs curriculum learning to train expert trajectories with more diverse supervision signals from the original graph. These expert trajectories are then used to optimize the condensed graph through an expanding window matching strategy, which gradually expands the matching range to balance the influence of easy and difficult nodes. Additionally, a knowledge embedding extractor is designed to further extract information from the expert trajectories. The proposed method is evaluated on various datasets and GNN architectures, demonstrating superior performance compared to state-of-the-art methods. GEOM achieves lossless graph condensation for 20 out of 35 cross-architecture experiments, opening up possibilities for broader real-world applications of graph condensation. The paper also includes theoretical analysis and ablation studies to validate the effectiveness of the proposed approach.