BLISS: Efficient GNN Training with Adaptive Node Sampling

This paper addresses the computational bottleneck of training Graph Neural Networks (GNNs) on large graphs. The core contribution is BLISS, a novel Bandit Layer Importance Sampling Strategy. By using multi-armed bandits, BLISS dynamically selects the most informative nodes at each layer, adapting to evolving node importance. This adaptive approach distinguishes it from static sampling methods and promises improved performance and efficiency. The integration with GCNs and GATs demonstrates its versatility.

• •BLISS introduces a novel bandit-based sampling strategy for GNN training.
• •It dynamically selects informative nodes, adapting to node importance.
• •BLISS integrates with GCNs and GATs, demonstrating versatility.
• •Experiments show BLISS maintains or exceeds full-batch training accuracy.