EUVIMEDEUVIMED entdecken
Crossref

GHMCL: A diffusion-based generative hypergraph framework with multi-scale contrastive learning

Weijun Gao, Maotang Su, Zhuyu Chen

International Journal of Modern Physics C · 22.08.2026

DOI: 10.1142/s0129183127501294 ↗

Kurz gefasst

Fachartikel in der Zeitschrift International Journal of Modern Physics C, online erschienen am 22.08.2026. Verfasst von Weijun Gao und 2 weiteren Autor:innen. Die Arbeit nennt 35 Literaturangaben.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Multimodal emotion recognition in conversation (MERC) requires models to capture high-order semantic relationships across modalities while staying robust to noise and structural uncertainty. Existing hypergraph methods usually depend on static or manually designed hyperedge structures, failing to model the complex, dynamically evolving dependencies in multimodal data. To tackle this issue, we propose the GHMCL framework. It generates dynamic hypergraphs through diffusion processes and integrates multi-scale contrastive learning to synergistically optimize high-order structure modeling and discriminative feature learning. Concretely, we first build a diffusion-based hypergraph generation module. Via a stepwise denoising diffusion process, it adaptively generates and optimizes hyperedge structures, capturing flexible, globally consistent high-order semantic relationships and explicitly modeling structural uncertainty. On this basis, we design a multi-scale contrastive learning mechanism, covering local contrast between nodes and hyperedges, global contrast between same and different-type semantics and an uncertainty confidence weighting mechanism derived from hyperedge weights. This boosts the model’s robustness to noisy and ambiguous samples, strengthening feature representations with local structural consistency and global category discriminability. Experiments were conducted on the public datasets MELD and IEMOCAP. The results show that GHMCL outperforms mainstream baseline models in both F1-score and accuracy, thus verifying the effectiveness of the proposed GHMCL framework in multimodal emotion recognition tasks.

Zitieren

APA

Gao, W., Su, M., & Chen, Z. (2026). GHMCL: A diffusion-based generative hypergraph framework with multi-scale contrastive learning. International Journal of Modern Physics C. https://doi.org/10.1142/s0129183127501294

Vancouver

Gao W, Su M, Chen Z. GHMCL: A diffusion-based generative hypergraph framework with multi-scale contrastive learning. International Journal of Modern Physics C. 2026. doi:10.1142/s0129183127501294

BibTeX
@article{Gao2026,
  author = {Gao, Weijun and Su, Maotang and Chen, Zhuyu},
  title = {GHMCL: A diffusion-based generative hypergraph framework with multi-scale contrastive learning},
  journal = {International Journal of Modern Physics C},
  year = {2026},
  doi = {10.1142/s0129183127501294}
}
Kostenlos starten →