Contrastive Pre-Training vs. Supervised Fine-Tuning in Few-Shot Node Classification on Low-Dimensional Graphs

This report synthesises findings from 14 peer-reviewed papers addressing the following research question: How do contrastive pre-training objectives for graph neural networks compare to supervised fine-tuning in few-shot node classification tasks on low-dimensional attribute graphs. Convolutional neural networks (CNNs) have achieved great success on grid-like data such as images, but face tremendous challenges in learning from more generic data such as graphs. In CNNs, the trainable local filters enable the automatic extraction of high-level features. 11 claims were extracted from source literature; 11 were independently verified against retrieved documents. An automated multi-reviewer quality assessment produced a score of 9.0/10. This report is a machine-generated literature synthesis and does not constitute original research.

Research goal: How do contrastive pre-training objectives for graph neural networks compare to supervised fine-tuning in few-shot node classification tasks on low-dimensional attribute graphs?

Autonomous literature synthesis. Automated review score: 9.0/10. Full text and citation available at Assignee Research.