Scalable Graph Neural Network Method for Efficient Learning on Large-Scale Billion-Node Graphs

Description of Generated Datasets and Figures Figure 7 – Accuracy Comparison:This dataset presents node-classification accuracy (%) across increasing graph sizes (1M to 1.2B nodes). It compares the proposed scalable GNN with a baseline model and includes a method-wise comparison on the 1.2B-node graph. The data demonstrate that the proposed framework consistently achieves higher accuracy, validating structural fidelity preservation at extreme scale. Figure 8 – Training Time:This table records end-to-end training time (in hours) for baseline and proposed models across graph sizes. The values highlight substantial reductions in training duration, evidencing the effectiveness of multi-level partitioning and distributed message passing. Figure 9 – GPU Memory Consumption:This dataset captures per-GPU memory usage (GB) for both models. It shows significant memory savings achieved by sparse representations, subgraph sampling, and activation checkpointing, enabling billion-node training on limited hardware. Figure 10 – Communication Overhead:Communication time per epoch (seconds) is reported for different scales. The reduced overhead in the proposed model reflects minimized boundary exchanges and compressed messaging in distributed execution. Figure 11 – Convergence Behavior:This table lists the number of epochs required to reach convergence. The proposed framework converges faster at all scales, indicating stable gradients and efficient propagation dynamics. Figure 12 – Scalability Efficiency:Strong-scaling efficiency (%) across GPU counts is provided for the proposed framework and DistDGL. The data demonstrate superior parallel efficiency, confirming balanced workload distribution and reduced synchronization cost. Figure 13 – Throughput Performance:Throughput in million edges per second is reported for baseline and proposed models. The proposed framework achieves markedly higher throughput, reflecting optimized kernels, reduced memory traffic, and efficient communication. These datasets collectively substantiate the claims of improved accuracy, efficiency, memory economy, communication reduction, faster convergence, superior scalability, and higher throughput of the proposed scalable GNN framework.