HyperTrust Enhances Hypergraph Neural Networks Against Label Noise

Hypergraph Neural Networks (HGNNs) are powerful for modeling complex higher-order relationships but are highly susceptible to label noise, which is common in real-world datasets. While methods exist for learning with label noise in general machine learning and graph learning, their application to hypergraphs has been largely unexplored. This research systematically investigates hypergraph node classification under label noise. The study first adapts and evaluates existing robust learning strategies for hypergraphs, revealing their limitations. Building on these insights, the authors propose HyperTrust, a novel framework. HyperTrust begins by estimating the trustworthiness of hyperedges using a pretraining-based, entropy-aware strategy. This initial step helps identify potentially reliable connections within the hypergraph. Subsequently, HyperTrust employs two key modules: HyperedgeBoost, which enhances reliable supervision by linking unlabeled nodes to trustworthy hyperedges, and HyperedgePrune, which actively suppresses noisy propagation by removing untrustworthy node-hyperedge incidences. These modules work collaboratively to refine the hypergraph structure and generate more accurate predictions, demonstrating superior effectiveness and robustness across various datasets and noise settings.