Kai Yin
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Research Interests
Kai Yin has a total of 12 co-authors. Here are the top co-authors by number of shared publications:
- Hao Huang
Wuhan University
2 shared publications - Yuxiang Zhao
Wuhan University
2 shared publications Hao Chi
Hubei Provincial Ecological Environment Information Center
1 shared publicationRuiting Wang
School of Computer Science, Wuhan University, Wuhan, China
1 shared publication- Shihong Yao
School of Geography and Information Engineering, China University of Geosciences, Wuhan, China
1 shared publication - Tao Wang
Hubei Key Laboratory of Digital Education, Central China Normal University, Wuhan, China
1 shared publication
Latest Publications
- 1 November 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Consumer Electronics
- Vol. 72 (4) , 11231-11241
Generative Structural Refinement via Importance-Aware Hypergraph Learning for Robust Higher-Order Representations
Hypergraph modeling of higher-order interactions is essential in fields such as emerging consumer electronics ecosystems. However, robust hypergraph representation learning remains challenging due to the complexity of multi-way relationships and the prevalence of structural noise. Although conventional Hypergraph Neural Networks (HGNNs) often incorporate attention mechanisms, they typically rely on local feature aggregation over the observed incidence structure. This dependence can limit their ability to learn noise-invariant structural representations, rendering the models vulnerable to structural perturbations and distribution shifts. To address these limitations, we propose Importance-aware Hypergraph Contrastive Learning (IHCL), a robust, generative-driven framework that integrates importance-guided structural perturbations with contrastive learning. IHCL quantifies hyperedge importance by combining four complementary signals: hyperedge degree, intra-edge feature variance, structural centrality, and feature centrality. Guided by these importance scores, IHCL progressively masks noise-sensitive hyperedges and enforces representation consistency between the original and generatively perturbed hypergraphs through contrastive learning. By effectively filtering out noisy connections while preserving discriminative higher-order patterns, IHCL substantially improves robustness of complex interaction systems. Theoretically, we show that the proposed importance-aware masking reduces expected representation shift and yields a tighter InfoNCE-based mutual-information lower bound than random masking under the same budget. Extensive experiments on eight topologically complex stress-test datasets demonstrate that IHCL outperforms state-of-the-art baselines, exhibiting superior stability under structural noise and diverse data distributions.
- 1 November 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Consumer Electronics
- Vol. 72 (4) , 11231-11241
Generative Structural Refinement via Importance-Aware Hypergraph Learning for Robust Higher-Order Representations
Hypergraph modeling of higher-order interactions is essential in fields such as emerging consumer electronics ecosystems. However, robust hypergraph representation learning remains challenging due to the complexity of multi-way relationships and the prevalence of structural noise. Although conventional Hypergraph Neural Networks (HGNNs) often incorporate attention mechanisms, they typically rely on local feature aggregation over the observed incidence structure. This dependence can limit their ability to learn noise-invariant structural representations, rendering the models vulnerable to structural perturbations and distribution shifts. To address these limitations, we propose Importance-aware Hypergraph Contrastive Learning (IHCL), a robust, generative-driven framework that integrates importance-guided structural perturbations with contrastive learning. IHCL quantifies hyperedge importance by combining four complementary signals: hyperedge degree, intra-edge feature variance, structural centrality, and feature centrality. Guided by these importance scores, IHCL progressively masks noise-sensitive hyperedges and enforces representation consistency between the original and generatively perturbed hypergraphs through contrastive learning. By effectively filtering out noisy connections while preserving discriminative higher-order patterns, IHCL substantially improves robustness of complex interaction systems. Theoretically, we show that the proposed importance-aware masking reduces expected representation shift and yields a tighter InfoNCE-based mutual-information lower bound than random masking under the same budget. Extensive experiments on eight topologically complex stress-test datasets demonstrate that IHCL outperforms state-of-the-art baselines, exhibiting superior stability under structural noise and diverse data distributions.
- Open Access
- 12 April 2026
- Published by Association for Computing Machinery (ACM)
- p. 3102-3110
DeepUL: Deep Unlearning via Model Sparsity
- 12 April 2026
- Published by Association for Computing Machinery (ACM)
- p. 3102-3110
Open AccessDeepUL: Deep Unlearning via Model Sparsity