STA-GANN: A Valid and Generalizable Spatio-Temporal Kriging Approach Conference: Proceedings of the 34th ACM International Conference on Information and Knowledge Management; November 10–14, 2025; Seoul, Republic of Korea Proceedings of the 34th ACM International Conference on Information and Knowle
Yujie Li, Zezhi Shao, Chengqing Yu, Tangwen Qian Zhao Zhang Yifan Du, Shaoming He, Fei Wang, Yongjun Xu · 2025
Read STA-GANN: A Valid and Generalizable Spatio-Temporal Kriging Approach Conference: Proceedings of the 34th ACM International Conference on Information and Knowledge Management; November 10–14, 2025; Seoul, Republic of Korea Proceedings of the 34th ACM International Conference on Information and Knowle by Yujie Li, Zezhi Shao, Chengqing Yu, Tangwen Qian Zhao Zhang Yifan Du, Shaoming He, Fei Wang, Yongjun Xu (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2508-16161.txt.
Introduction
Li and co-authors present STA-GANN, a method for filling in missing readings across space and time when sensors fail or cannot be reached. Such gaps are common in sensor networks, and guessing the missing stretches well is hard. Their 2025 paper builds a graph based neural network designed to stay valid and to work on sensors it never saw in training.
Three parts drive the design. One senses and adjusts for shifts in timestamps. A second rebuilds the map of spatial links from fresh data and side details. A third uses adversarial transfer learning to carry patterns to new sensors. Tests across nine datasets from four fields, plus theory checks, showed top performance.
Catalog record
| CREATOR | Yujie Li, Zezhi Shao, Chengqing Yu, Tangwen Qian Zhao Zhang Yifan Du, Shaoming He, Fei Wang, Yongjun Xu |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 45 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2508-16161
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
STA-GANN: A Valid and Generalizable Spatio-Temporal Kriging Approach Conference: Proceedings of the 34th ACM International Conference on Information and Knowledge Management; November 10–14, 2025; Seoul, Republic of Korea Proceedings of the 34th ACM International Conference on Information and Knowledge Management (CIKM ’25), November 10–14, 2025, Seoul, Republic of Korea DOI: 10.1145/3746252.3761045 ISBN: 979-8-4007-2040-6/2025/11 CCS: Information systems Data mining Yujie Li, Zezhi Shao, Chengqing Yu, Tangwen Qian Zhao Zhang Yifan Du, Shaoming He, Fei Wang, Yongjun Xu Abstract . Spatio-temporal tasks often encounter incomplete data arising from missing or inaccessible sensors, making spatio-temporal kriging crucial for inferring the completely missing temporal information. However, current models struggle with ensuring the validity and generalizability of inferred spatio-temporal patterns, especially in capturing dynamic spatial dependencies and temporal shifts, and optimizing the generalizability of unknown sensors. To overcome these limitations, we propose Spatio-Temporal Aware Graph Adversarial Neural Network (STA-GANN) , a novel GNN-based kriging framework that improves spatio-temporal pattern validity and generalization. STA-GANN integrates (i) Decoupled Phase Module that senses and adjusts for timestamp shifts. (ii) Dynamic Data-Driven Metadata Graph Modeling to update spatial relationships using temporal data and metadata; (iii) An adversarial transfer learning strategy to ensure generalizability. Extensive validation across nine datasets from four fields and theoretical evidence both demonstrate the superior performance of STA-GANN. † † cc-license: by 1. Introduction Spatio-temporal tasks are crucial in domains such as energy, transportation ( Zheng et al. 2020 ; Shao et al. 2022a ; Shao et al. 2022c ) , and meteorology ( Bi et al. 2023 ) , yet sensor unavailability caused by deployment costs, equipment failures often leads to severe data gaps. To mitigate it, spatio-temporal kriging ( Wu et al. 2021b ) has become a key technique for inferring completely missing time series through temporal and spatial dependencies of available sensors. The primary challenge of spatio-temporal kriging lies in the absence of time series data for certain unknown sensors or nodes. Unlike conventional time series imputation, which utilizes historical and future data to fill in missing values within the same series, spatio-temporal kriging relies on temporal information from other sensors and their spatial relationships to make inferences. As a result, the model must capture spatio-temporal patterns ( Li et al. 2025 ) in both a valid and generalizable manner. This involves mining meaningful patterns from available data and ensuring these patterns can be transferred to sensors or nodes without prior observations. Figure 1. Challenges of Spatio-Temporal Kriging. (A) Predefined graphs may contain unpredictable errors because some edge relations are incorrectly established or ignored. (B) Timestamp shift refers to delays in the transmission of temporal information caused by inevitable factors such as distance and upstream-downstream relationships. (C) The captured overfitted patterns may only be suitable for specific sensors and cannot be generalized. Existing methods often struggle to the dual requirements of validity and generalization in spatio-temporal kriging. Traditional approaches, such as Kriging ( Appleby et al. 2020 ) and matrix decomposition ( Bahadori et al. 2014 ) , typically underperform compared to deep learning models ( Xu et al. 2021 ) in real-world scenarios. IGNNK ( Wu et al. 2021b ) is the first to adopt graph neural networks ( Kipf and Welling 2017 ; Gilmer et al. 2017 ) for Kriging, establishing the foundation for subsequent research. SATCN ( Wu et al. 2021c ) , IAGCN ( Wei et al. 2024 ) , and INCREASE ( Zheng et al. 2023 ) all aim to improve performance by extending spatial relationships, but differ in focus: INCREASE leverages scarce metadata including functional areas and points of interest (POIs), while SATCN and IAGCN focus on optimizing the predefined graph ( Yu et al. 2018 ; Li et al. 2018 ) or extracting features from it. DualSTN ( Hu et al. 2023 ) instead targets temporal patterns by disentangling long- and short-term dependencies. However, these methods still face challenges in capturing valid and generalizable spatio-temporal patterns: • For the validity of spatial patterns, errors in predefined graphs lead to error accumulation during message passing, and current work fails to capture the dynamics of spatial relationships. ( Wu et al. 2019 ; Yu et al. 2023 ; Li et al. 2023 ) . • For the validity of temporal patterns ( Shih et al. 2019 ; Shao et al. 2023 ) , timestamp shift poses a unique challenge for kriging, which is distinct from other temporal issues like long-term dependence. • For the generalization of spatio-temporal patterns to unknown nodes, Label propagation methods for graph-based semi-supervised learning only guarantee convergence ( Iscen et al. 2019 ) , but overlooks generalization to unknown sensor patterns. To address the key challenges of spatio-temporal kriging in terms of validity and generalization, we decompose them into three fundamental challenges and propose corresponding solutions and innovations, as illustrated in Figure 1 : 1. How to ensure the spatial relationship is appropriate? Spatial relationships are crucial for spatio-temporal kriging. Predefined graphs ( Yu et al. 2018 ; Li et al. 2018 ) , as shown in Figure 1 (A), misrepresent sensor spatial patterns and lack adaptability to environmental changes. This diminishes sensor interactions and causes error accumulation in graph learning, ultimately degrading kriging performance. OUR APPROACH: We propose Dynamic Data-Driven Metadata Graph Modeling (D3MGM) for kriging to capture dynamic spatial dependencies among sensors. This framework leverages temporal data and metadata ( Li et al. 2024 ; Liu et al. 2024 ) , enabling the model to utilize universal information such as coordinates and timestamps for constructing accurate spatial relationships. 2. How to learn effective temporal patterns? Timestamp shift in Figure 1 (B) ( Niu et al. 2023 ; Li et al. 2024 ) refers to the signal transmission delay between sensors. This results in the same information being described at different timestamps across sensors, leading to misalignment with target values and degraded metrics. Moreover, this is a unique problem for spatio-temporal kriging, as forecasting relies on historical data and imputation uses contextual information. OUR APPROACH: We introduce Decoupled Phase Module (DPM) to sense and compensate for timestamp shifts. By adjusting the phase of unknown sensors in the frequency domain ( Zhou et al. 2022 ; Yi et al. 2024 ) using decoupled series ( Wu et al. 2021a ) , we leverage the Fourier Transform’s time-shift property ( Yang et al. 2020 ) to align timestamps when specific time series information emerges. 3. How to ensure that learned spatio-temporal patterns are generalized? During training, unknown sensors are absent, and missing nodes are simulated by randomly masking known sensors. This raises a critical question: how to ensure the model’s generalization to truly unknown sensors and new spatial relationships? OUR APPROACH: Despite the importance of inductive GNNs for spatio-temporal kriging, it is still difficult to ensure the generalization of spatio-temporal patterns during label propagation. ( Iscen et al. 2019 ) . To address this, we apply adversarial transfer learning ( Ganin et al. 2016 ) to enhance the generalizability ( Du et al. 2021 ; Lu et al. 2022 ) . We treat known nodes as the source domain and unknown nodes as the target domain. and introduce discriminators which motivates the model to distinguish between known and unknown sensors. If the model successfully distinguishes