A Study on Active Learning for Graphs
Pratheeksha Nair, Zhi Wen
Abstract
This is a comprehensive study of active learning methods on real-world graph datasets. Active Learning for interconnected data has been studied over the years and has gained increasing importance in recent times, especially due to its applications in tasks where labeling is laborious and requires human experts, such as drug discovery and protein-protein interaction prediction. In this work, various active learning strategies are compared across 6 real-world datasets from different domains for the task of node classification. The main goal of this evaluation is to benchmark a range of active learning strategies against state-of-the-art and identify the ones that consistently perform well. We also propose a simple strategy for selecting nodes for training the node classifier, and our experiments show promising results of this strategy.