Dual-Level Adaptive Correction for Subgraph-Based GNN Training with Efficient Strategy Search

Graph neural networks have achieved remarkable performance across a wide range of complex graph-based tasks. While subgraph sampling (SS) methods substantially improve per-epoch efficiency for large graphs, they introduce higher gradient variance, which hinders convergence and reduces accuracy. Moreover, applying corrections to reduce SS variance impact faces the issue of diminishing returns, thereby limiting training efficiency. To address these challenges, we propose ECHO+, a novel dual-level correction framework designed to accelerate training while maintaining accuracy comparable to that of node sampling. ECHO+ employs a lightweight, variance-guided strategy search during preprocessing to reduce SS variance. During training, it operates on two levels: a coarse level that adaptively schedules correction to reduce residual SS variance, and a fine level that leverages a batch-loss driven early stopping mechanism to improve overall training efficiency. Experimental results reveal that ECHO+ achieves rapid convergence with high accuracy, delivering a speedup of up to 12.4x over the node sampling baseline while maintaining comparable performance. Furthermore, ECHO+ outperforms existing SS baselines, achieving up to 4.3x faster convergence.