Coarse-to-Fine Label Distribution Learning for High-Cardinality Survey Simulation
International Journal of Web Services Research · 09.10.2026 · 23 (1) · 1-19
DOI: 10.4018/ijwsr.424308 ↗Kurz gefasst
Fachartikel in der Zeitschrift International Journal of Web Services Research, erschienen am 09.10.2026. Verfasst von Jianhao Li und 4 weiteren Autor:innen. Die Arbeit nennt 30 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 3.0).
Abstract (Originalsprache)
Modeling fine-grained survey distributions for high-cardinality large language model prediction is challenging under mixed-cardinality supervision. Because low-cardinality surveys dominate training, standard label distribution learning (LDL) develops positional bias toward early response regions and fails to generalize probability allocation to extended spaces. This bias collapses probability mass into low-index regions, thereby fragmenting high-cardinality distributions. To address this, the authors propose hierarchical LDL, a coarse-to-fine framework using macroscopic anchors to calibrate fine-grained predictions. By jointly optimizing hierarchical distributions, hierarchical LDL stabilizes probability allocation across sparsely supervised high-cardinality scales and improves structural alignment.
Zitieren
APA
Li, J., Liu, H., Meng, Z., Qiu, C., & Gu, J. (2026). Coarse-to-Fine Label Distribution Learning for High-Cardinality Survey Simulation. International Journal of Web Services Research, 23(1), 1-19. https://doi.org/10.4018/ijwsr.424308
Vancouver
Li J, Liu H, Meng Z, Qiu C, Gu J. Coarse-to-Fine Label Distribution Learning for High-Cardinality Survey Simulation. International Journal of Web Services Research. 2026;23(1):1-19. doi:10.4018/ijwsr.424308
BibTeX
@article{Li2026,
author = {Li, Jianhao and Liu, Haijiang and Meng, Zixiang and Qiu, Chen and Gu, Jinguang},
title = {Coarse-to-Fine Label Distribution Learning for High-Cardinality Survey Simulation},
journal = {International Journal of Web Services Research},
year = {2026},
volume = {23},
number = {1},
pages = {1--19},
doi = {10.4018/ijwsr.424308}
}
EUVIMED