GATTA: Enhancing Active Learning on Graphs with Test-Time Augmentation
A novel framework named GATTA (Graph Active Learning with Test-Time Augmentation) has been developed to enhance active learning for graph-structured data. While test-time augmentation (TTA) has proven beneficial in computer vision for strengthening model resilience and estimating uncertainty, its use in graph contexts has been scarce. GATTA consolidates predictions from various augmented views to yield more trustworthy uncertainty assessments. To address the challenges of label-preserving graph augmentations, it employs a consistency-based filtering method that eliminates augmented views with unreliable predictions. The framework underwent thorough evaluation across various graph datasets, GNN architectures, and acquisition strategies. Findings show that basic uncertainty-based techniques like Entropy and Least Confidence gain the most from TTA, achieving results comparable to more complex and resource-intensive methods. The paper can be accessed on arXiv under identifier 2608.15084.
Key facts
- GATTA is a framework for active learning on graphs using test-time augmentation.
- It aggregates predictions across multiple augmented views for better uncertainty estimates.
- A consistency-based filtering mechanism discards unreliable augmented views.
- Evaluated on multiple graph datasets, GNN architectures, and acquisition strategies.
- Simple uncertainty methods like Entropy and Least Confidence benefit most from TTA.
- GATTA achieves performance competitive with more complex approaches.
- Paper available on arXiv with ID 2608.15084.
- Test-time augmentation has been effective in computer vision but underexplored for graphs.
Entities
Institutions
- arXiv