Nov 2026· Proceedings of the 35th ACM International Conference on Information and Knowledge Management· 0 citations· 52 references
Abstract
How do we classify nodes accurately on heterophilic graphs when training labels are infected by neighbors' majority classes? In many real-world applications, annotators lack sufficient evidence for each node in isolation and thus rely on its neighborhood context, causing labels to drift toward the neighborhood majority. We call this bias structural label infection, a graph-dependent corruption that makes supervision appear artificially homophilic even though the underlying graph remains heterophilic. This setting is particularly challenging because message passing quickly reinforces infected labels, making corrupted nodes look confident and rendering common loss- or agreement-based denoising methods unreliable. Structural label infection affects both hop selection and label correction. Because corrupted labels are aligned with neighborhood majorities, optimizing hop weights with these labels gives fake-homophilic neighborhoods an artificial advantage. This is especially harmful on heterophilic graphs, where useful propagation is not restricted to immediate neighbors. Moreover, predictions are produced by a model trained on infected labels, so directly copying them passes this infection to the next training round. We propose Recover (Robust nodE Classification under structural label nOise Via multi-score hop routing and prototypE Relabeling), a robust node classification framework that severs the feedback loop between infected supervision and structural propagation. Instead of allowing corrupted labels to dictate neighborhood aggregation, Recover secures robust representations by dynamically routing propagation through proxies based on static features and averaged task confidence. Subsequently, it directly rectifies the corrupted supervision by projecting these stabilized embeddings against reliable class prototypes. Extensive experiments demonstrate that Recover achieves up to 29.41 percentage-point improvement in accuracy under infected-label settings.
Sipu is proposed, an accurate graph-based PU learning method for heterophilic graphs that mitigates negative transfer from heterophilic neighbors while preserving higher-order homophilic information through positive paths.
Shihyun Park, J. Kim, U. Kang· Proceedings of the 35th ACM...· 0 citations
Graph Convolutional Networks (GCNs) have demonstrated exceptional performance on graph-structured data. However, their success heavily relies on the availability of clean labels, an assumption that often fails in real-world scenarios. The learnability of GCNs is intrinsically linked to the neighborhood structure of nod...
The rapid proliferation of social media has accelerated multimodal misinformation, where false claims paired with manipulated images make detection substantially more challenging. Existing multimodal fake news detection methods typically predict each sample independently, ignoring inter-sample relational evidence. Whil...
Ze-Jiang He, Luo Yun, Fa-Ying Wu et al.· Proceedings of the 34th ACM...· 0 citations
How does noisy supervision change the representations learned by a neural network? Training accuracy measures agreement with training labels but provides little information about representation geometry. We use persistent homology to track loop structure across distance thresholds in neural activations. Experiments spa...
Zhong-Tian Sun, Fan Mo, Prayag Tiwari et al.· 0 citations
Graph-level out-of-distribution (OOD) generalization is critical for deploying graph neural networks in real-world applications where distribution shifts are inevitable. Existing approaches typically require class labels during representation learning to decide which substructures are predictive, limiting their applica...
Tian-Jun Yao, Cheng-Hu Xi, Xin-Yang Li et al.· 0 citations
Large-scale datasets inevitably contain noisy labels, compelling deep neural networks to overfit and degrade in performance. Existing learning with noisy labels (LNL) approaches partially alleviate this through sample selection. However, selection processes themselves inherently rely on models trained on noisy data, wh...
Sen-Yu Hou, Gao-Xia Jiang, Xin-Yi Zheng et al.· IEEE Transactions on Pattern...· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.