Jintao Gu
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Research Subjects
Jintao Gu has a total of 20 co-authors. Here are the top co-authors by number of shared publications:
- K. Max Zhang
Cornell University
2 shared publications Ali Amadeh
Cornell University
1 shared publication- Bowen Xiao
School of Economics and Management, Beihang University, Xueyuan Road, Haidian District, Beijing 100191, China
1 shared publication - Ce Jia
School of Economics and Finance, Hohai University, Changzhou 213200, China
1 shared publication Chuan Liao
1 shared publicationHuijuan Wang
1 shared publication
Latest Publications
- Open Access
- 28 September 2026
- Published by IOP Publishing in Environmental Research Letters
- Vol. 21 (19) , 194003
The interdependence between cross-validation and hyperparameter tuning reveals limits of performance metrics and interpretation in random-forest PM 2 . 5 models
Random forest (RF) models for fine particulate matter (PM 2.5 ) prediction inform exposure assessment and air-quality management. These models are typically evaluated using spatial or random cross-validation (CV), which often yields high R 2 . However, how the choice of CV strategy interacts with hyperparameter tuning (e.g. max_features) has received little attention. Using daily PM 2.5 data from 157 monitoring stations in northern China, we show that Spatial-CV and Temporal-CV produce substantially different performance ( R 2 of 0.85 vs 0.37) and, more importantly, lead to opposing optimal hyperparameter settings. The clearest case is max_features: Spatial-CV performs best when almost all features are considered at each split, while Temporal-CV performs best with only a few (20 vs 3 of the 22 features). The performance gap and the hyperparameter reversal between Spatial-CV and Temporal-CV both reflect spatial interpolation rather than learned physical relationships. A diagnostic Date-ID model built from date and coordinate identifiers alone (no physical predictors) reproduced both patterns and reached the highest Spatial-CV R 2 (0.95); conversely, reducing station density from 157 to 20 stations progressively weakened both. For the current literature, the implication is direct: a high RF model score (e.g. R 2 ) cannot certify what a model has learned or how far it generalizes. A predictor’s assessed value and the model’s overall interpretation depend on the data structure, evaluation, and hyperparameters, and can reverse when these change. Satellite aerosol optical depth, for instance, appears unnecessary under Spatial-CV but helps under Temporal-CV, and feature-importance rankings shift between the optimal hyperparameter configurations of Spatial- and Temporal-CV. This indicates a systematic mismatch between what a CV metric measures and what is claimed from it in the literature. We recommend tuning hyperparameters separately for each CV strategy, reporting performance across station densities, and comparing against a date-and-coordinate baseline before model outputs are used for exposure assessment or monitoring-network design.
- 28 September 2026
- Published by IOP Publishing in Environmental Research Letters
- Vol. 21 (19) , 194003
Open AccessThe interdependence between cross-validation and hyperparameter tuning reveals limits of performance metrics and interpretation in random-forest PM 2 . 5 models
Random forest (RF) models for fine particulate matter (PM 2.5 ) prediction inform exposure assessment and air-quality management. These models are typically evaluated using spatial or random cross-validation (CV), which often yields high R 2 . However, how the choice of CV strategy interacts with hyperparameter tuning (e.g. max_features) has received little attention. Using daily PM 2.5 data from 157 monitoring stations in northern China, we show that Spatial-CV and Temporal-CV produce substantially different performance ( R 2 of 0.85 vs 0.37) and, more importantly, lead to opposing optimal hyperparameter settings. The clearest case is max_features: Spatial-CV performs best when almost all features are considered at each split, while Temporal-CV performs best with only a few (20 vs 3 of the 22 features). The performance gap and the hyperparameter reversal between Spatial-CV and Temporal-CV both reflect spatial interpolation rather than learned physical relationships. A diagnostic Date-ID model built from date and coordinate identifiers alone (no physical predictors) reproduced both patterns and reached the highest Spatial-CV R 2 (0.95); conversely, reducing station density from 157 to 20 stations progressively weakened both. For the current literature, the implication is direct: a high RF model score (e.g. R 2 ) cannot certify what a model has learned or how far it generalizes. A predictor’s assessed value and the model’s overall interpretation depend on the data structure, evaluation, and hyperparameters, and can reverse when these change. Satellite aerosol optical depth, for instance, appears unnecessary under Spatial-CV but helps under Temporal-CV, and feature-importance rankings shift between the optimal hyperparameter configurations of Spatial- and Temporal-CV. This indicates a systematic mismatch between what a CV metric measures and what is claimed from it in the literature. We recommend tuning hyperparameters separately for each CV strategy, reporting performance across station densities, and comparing against a date-and-coordinate baseline before model outputs are used for exposure assessment or monitoring-network design.
- 22 January 2026
- Published by Elsevier in Energy Policy
- Vol. 211
How should China exit clean heating subsidies? Policy design for the post-subsidy transition
- 22 January 2026
- Published by Elsevier in Energy Policy
- Vol. 211
How should China exit clean heating subsidies? Policy design for the post-subsidy transition
- 9 August 2025
- Published by Elsevier in Environmental Research
- Vol. 285 (Pt 4) , 122540
Integrated screening techniques reveal insight into hyperlocal non-traffic emission sources
- 9 August 2025
- Published by Elsevier in Environmental Research
- Vol. 285 (Pt 4) , 122540
Integrated screening techniques reveal insight into hyperlocal non-traffic emission sources
- 13 October 2024
- Published by Elsevier in Sustainable Cities and Society
- Vol. 116
Unraveling the global economic and mortality effects of rising urban heat island intensity
- 13 October 2024
- Published by Elsevier in Sustainable Cities and Society
- Vol. 116
Unraveling the global economic and mortality effects of rising urban heat island intensity
- 1 December 2023
- Published by SPIE-Intl Soc Optical Eng
- Vol. 12940, 863-867
Lightweight facial attribute editing with separable latent vector
There are two main challenges faced by generative adversarial networks (GANs) for facial attribute editing. The first challenge is how to perform targeted attribute editing in a controllable manner during the editing process, preserving the variability of relevant attributes while ensuring the invariance of irrelevant attributes. The second challenge is the high computational resource requirements of GANs, making them demanding on hardware performance for almost all image editing GANs. To address these challenges, we propose our SLGAN model based on StyleGAN2, which incorporates separable loss and knowledge distillation methods. Experimental results demonstrate that our proposed model achieves promising performance on relevant tasks.
- 1 December 2023
- Published by SPIE-Intl Soc Optical Eng
- Vol. 12940, 863-867
Lightweight facial attribute editing with separable latent vector
There are two main challenges faced by generative adversarial networks (GANs) for facial attribute editing. The first challenge is how to perform targeted attribute editing in a controllable manner during the editing process, preserving the variability of relevant attributes while ensuring the invariance of irrelevant attributes. The second challenge is the high computational resource requirements of GANs, making them demanding on hardware performance for almost all image editing GANs. To address these challenges, we propose our SLGAN model based on StyleGAN2, which incorporates separable loss and knowledge distillation methods. Experimental results demonstrate that our proposed model achieves promising performance on relevant tasks.