Stochastic Mirror-Prox Optimisation of Group Distributionally Robust Multimodal Logistic Regression for Multilingual Fact-Checking Verdict Prediction
4 affiliations
- 1Faculty of Arts and Social Sciences, The University of Sydney, Sydney, NSW 2006, Australia
- 2School of Economics, Peking University, Beijing 100871, China
- 3China-ASEAN Institute of Statistics, Guangxi University of Finance and Economics, Nanning 530003, China
- 4Graduate School of Technology and Management, Kyung Hee University, Yongin-si 17104, Gyeonggi-do, Republic of Korea
Abstract
This paper asks whether group distributionally robust optimisation raises the worst-group macro-F1 of a block-masked multimodal logistic regression with availability indicators for multilingual fact-checking. Stochastic Mirror-Prox solves the resulting convex–concave saddle-point problem. The corpus pools 49,584 X-Fact and MediaEval records with 90 language labels. On the in-domain split, the robust solution lifts worst-group macro-F1 from 0.313 under pooled empirical risk minimisation to 0.401 (95% interval of the gain [0.025,0.134]). Overall macro-F1 drops by 0.040, and dispersion of balanced accuracy across languages falls from 0.166 to 0.131. The gain lies in X-Fact claim groups and the cost in MediaEval posts (macro-F1 0.887 to 0.809). Removing the availability indicators moves no worst-group or overall macro-F1 by more than 0.007. The worst-group difference is negative on unseen languages and not resolved on unseen sites or events. A Friedman test detects no difference between solvers in typical group performance. At equal numbers of sampled records, stochastic mirror descent reaches a smaller mean duality gap than Mirror-Prox at each measured mini-batch size from 2 to 128. On this instance, extrapolation does not repay its second call.
Related
- Distributionally robust linear regression through the lens of adversarial training
- T-ARC: Topology-Aware Randomized Clustering via Distributionally Robust Stochastic Block Models
- A Composite Bayesian Risk--Distributionally Robust Optimization Model for Stochastic Programming with Elliptical Distribution Ambiguity
- Particle-based Generalised Stochastic Optimisation
- A Stochastic EM Algorithm with Sampling-Importance Resampling for Missing Data in Regression with Nonlinear Predictors
BibTeX
@article{li2026stochastic,
title = {{Stochastic Mirror-Prox Optimisation of Group Distributionally Robust Multimodal Logistic Regression for Multilingual Fact-Checking Verdict Prediction}},
author = {Shuyun Li and Fan Si and Dongxu Mo and Chao He},
journal = {Mathematics},
year = {2026},
month = oct,
doi = {10.3390/math14193574},
url = {https://www.mdpi.com/2227-7390/14/19/3574},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Stochastic Mirror-Prox Optimisation of Group Distributionally Robust Multimodal Logistic Regression for Multilingual Fact-Checking Verdict Prediction
AU - Shuyun Li
AU - Fan Si
AU - Dongxu Mo
AU - Chao He
PY - 2026
DA - 2026/10/02
JO - Mathematics
DO - 10.3390/math14193574
UR - https://www.mdpi.com/2227-7390/14/19/3574
AB - This paper asks whether group distributionally robust optimisation raises the worst-group macro-F1 of a block-masked multimodal logistic regression with availability indicators for multilingual fact-checking. Stochastic Mirror-Prox solves the resulting convex–concave saddle-point problem. The corpus pools 49,584 X-Fact and MediaEval records with 90 language labels. On the in-domain split, the robust solution lifts worst-group macro-F1 from 0.313 under pooled empirical risk minimisation to 0.401 (95% interval of the gain [0.025,0.134]). Overall macro-F1 drops by 0.040, and dispersion of balanced accuracy across languages falls from 0.166 to 0.131. The gain lies in X-Fact claim groups and the cost in MediaEval posts (macro-F1 0.887 to 0.809). Removing the availability indicators moves no worst-group or overall macro-F1 by more than 0.007. The worst-group difference is negative on unseen languages and not resolved on unseen sites or events. A Friedman test detects no difference between solvers in typical group performance. At equal numbers of sampled records, stochastic mirror descent reaches a smaller mean duality gap than Mirror-Prox at each measured mini-batch size from 2 to 128. On this instance, extrapolation does not repay its second call.
SN - 2227-7390
ER -