ARTFEED — Contemporary Art Intelligence

GSSL Robustness to Real-World Noise in Biomedical Graphs

other · 2026-05-09

A new study evaluates Graph Self-Supervised Learning (GSSL) methods on text-driven biomedical knowledge graphs, addressing a gap in robustness research that previously focused on synthetic perturbations. The authors introduce NATD-GSSL, a framework combining automatic graph construction, refinement, and GSSL. Using a dual-graph protocol, they compare a noisy graph from MedMentions with a clean UMLS reference graph for unsupervised term typing. The work highlights challenges posed by real-world noise in automatically extracted graphs.

Key facts

  • First comprehensive evaluation of GSSL methods on text-driven graphs for unsupervised term typing.
  • Introduces NATD-GSSL framework.
  • Uses dual-graph protocol contrasting noisy MedMentions graph with clean UMLS reference graph.
  • Addresses real-world noise, not synthetic perturbations.
  • Focuses on biomedical domain.
  • Published on arXiv with ID 2605.05463.
  • Leverages NLP for automatic knowledge graph extraction.
  • Graph Self-Supervised Learning paradigm used.

Entities

Institutions

  • arXiv
  • MedMentions
  • Unified Medical Language System (UMLS)

Sources