Researchers at Georgia Tech Propose 'LABOR' (LAyer-neighBOR sampling), A New Sampling Algorithm-Based on Machine Learning

The de facto models for representation learning on graph-structured data are Graph Neural Networks (GNN). As a result, they have begun to be implemented in production systems. These models pass messages along the direction of the edges in the given graph with nonlinearities between different layers, updating the node embeddings iteratively. The computed node embeddings for l layers include details from the seed vertex's l-hop neighborhood. The GNN models must be trained on billion-scale graphs to be used in production.