Feature space variation-based active learning sample query strategy for graph deep learning

Xinlong Chen

doi:10.1016/j.eswa.2026.131165

Return

Feature space variation-based active learning sample query strategy for graph deep learning

2026-01-01

PRE

DOI：10.1016/j.eswa.2026.131165

Original

Original request for help

Save

Abstract

En 中文

Active learning (AL) has evolved diverse query strategies to select the most valuable samples for annotation to optimize deep learning (DL) models' efficiency. However, there is an undeniable challenge in that the feature space distribution of samples is reshaped by DL models, altering their underlying semantic information. The existing AL strategies often overlook this critical information, resulting in the unnecessary consumption of annotation resources and severely limiting the learning capabilities of DL models. Therefore, to tackle this challenge, this paper investigates the intricate relationship between the dynamic shifts in sample information and the learning potential of DL models from an innovative perspective, proposing Feature Space Variation-based (FSV) active learning sample query strategy. In specific scenarios, samples are categorized into Solid, Reversal, Potential, and Challenge data based on their importance. Concurrently, we have developed a targeted maximum entropy stability index to quantify the potential annotation value of samples effectively. Notably, this is the first sample query strategy in the AL domain based on feature space variation, offering a fresh perspective and approach. Finally, the proposed strategy is integrated with the GNN model for graph semi-supervised node classification and graph classification tasks across sixteen benchmarks. The results demonstrate that our strategy significantly boosts the DL model's performance and significantly conserves label resources, thoroughly substantiating the strategy's effectiveness.

Keywords:

Active learning

Query strategy

Graph neural networks

Semi-supervised learning

Node classification

Graph classification