(#96 / Sess. 1) Active Learning on Graphs via Meta Learning
Abstract
Active learning (AL) for semi-supervised node classification aims to reduce the number of labeled instances by selecting only the most informative nodes for labeling. The AL algorithms designed for other data types such as images and text do not perform well on graph-structured data. Although a few heuristics-based AL algorithms have been proposed for graphs, a principled approach is lacking. We propose MetAL, an AL algorithm that selects unlabeled items that directly improve the future performance of a graph neural network (GNN) model. We formulate the AL problem as a bilevel optimization problem. Based on recent work in meta-learning, we compute the meta-gradients to approximate the impact of unlabeled instances on the model uncertainty. We empirically demonstrate that MetAL outperforms existing AL algorithms.