SLeDGe: Semi-Supervised Learning on Data Streams with Graph Structure Learning

SLeDGe is a novel semi-supervised learning (SSL) method designed for data streams that integrates adaptive graph structure learning with a predictive model under strict memory and labeling constraints. It features distinct update strategies for maintaining compact labeled and unlabeled memories and encourages sparsity in the relational graph to enhance label supervision propagation. Evaluated across 12 datasets, SLeDGe demonstrates significant performance improvements, achieving average relative accuracy gains of 31.7% with only 0.1% labeled data and 14.8% with 1% labeled data, making it a valuable tool for practitioners dealing with evolving data streams.