Active learning in graph-based semi-supervised learning

Designed and evaluated an active learning framework for graph-based semi-supervised learning (GSSL) that intelligently selects the most informative data points to label. The approach combines uncertainty metrics (like least-confidence and entropy) with graph-theoretic centrality measures (e.g., betweenness, pagerank, and a novel clique-overlap centrality) to enhance diversity in selected instances. Experimental results on benchmark datasets demonstrate that least-confidence uncertainty with clique-overlap centrality significantly improves labeling efficiency and prediction accuracy. The framework was tested using multiple GSSL methods, including Label Propagation, Modified Adsorption, and Local & Global Consistency.

For details View Report