HyperTrust: A Robust Framework for Hypergraph Neural Networks under Label Noise
Researchers have unveiled a new framework called HyperTrust, aimed at strengthening hypergraph neural networks (HGNNs) against issues caused by label noise. Documented on arXiv as paper number 2608.04377, the study critically reviews the limitations of current approaches to learning amidst label noise and graph learning for hypergraphs. HyperTrust employs a pretraining strategy assessing hyperedge reliability through entropy and introduces a HyperedgeBoost module to enhance supervision by linking unlabeled nodes. The findings reveal the susceptibility of HGNNs to noisy labels, addressing a largely ignored area of noisy-label learning in hypergraph studies. The team seeks to improve hypergraph model dependability for complex relational tasks.
Key facts
- Hypergraph neural networks (HGNNs) are vulnerable to label noise.
- The paper presents a systematic study of hypergraph node classification under label noise.
- Representative LLN and GLN methods were adapted to hypergraphs and evaluated under a unified benchmark.
- Existing robust learning strategies show limitations for hypergraphs.
- HyperTrust is proposed as a new robust framework for hypergraphs.
- HyperTrust estimates hyperedge trustworthiness via a pretraining-based, entropy-aware strategy.
- HyperTrust incorporates the HyperedgeBoost module to enhance reliable supervision.
- The paper is available on arXiv with identifier 2608.04377.
Entities
Institutions
- arXiv