When Active Learning Meets Graph Similarity: Evidential Variance for Graph Selection

Abstract: Graph Similarity Learning (GSL) is pivotal in graph data mining, yet training effective models necessitates substantial labeled pairs, which incur prohibitive annotation costs. To address this, we introduce Active Learning (AL) into the GSL paradigm. However, directly transferring existing AL strategies is non-trivial due to two unique impediments: (1) the continuous regression nature of similarity prediction complicates standard uncertainty quantification, and (2) the paired-input structure requires evaluating a graph's informational value across its pairings rather than in isolation. To bridge this gap, we propose EVGS (Evidential Variance for Graph Selection), a novel AL framework tailored for GSL. EVGS leverages evidential deep learning to impose a prior over predictions, enabling disentangled uncertainty estimation. Crucially, we identify a ``gradient shrinkage'' pathology inherent to the data-scarce regime characteristic of AL cycles. We introduce a novel MSE-anchored regularizer to mitigate this issue, ensuring discriminative uncertainty estimation even with limited labels. Furthermore, to address the paired-input challenge, we propose a graph-centric selection criterion: uncertainty variance. This metric captures a graph's holistic informational value by measuring fluctuations in its epistemic uncertainty across diverse interactions. Extensive experiments on three benchmarks with two GSL backbones demonstrate that EVGS consistently outperforms established AL baselines.

Submission Type: Regular submission (no more than 12 pages of main content)

Assigned Action Editor: ~Tianshu_Yu2

Submission Number: 7562