TY - JOUR
T1 - Graph signal processing meets machine learning
T2 - Multi-scale spatial-temporal ensemble learning methodology for air quality forecasting
AU - Wang, Zicheng
AU - Chen, Liren
AU - Chen, Huayou
AU - Yang, Jingling
AU - Rehman, Naveed ur
N1 - Publisher Copyright:
© 2025 Elsevier Ltd
PY - 2025/10/1
Y1 - 2025/10/1
N2 - Reliable air quality forecasting is a challenging task, as we need to consider both complex temporal and spatial patterns. Following the philosophy of the well-established “decomposition and ensemble” modeling framework, this paper proposes a novel multi-scale spatial–temporal ensemble learning methodology based on time-varying graph mode decomposition (TVGMD) and machine learning (ML) techniques to jointly analyze the temporal dynamics and spatial interactions of multi-station air quality at multiple timescales, aiming to achieve more accurate air quality predictions. The suggested methodology comprises four multi-scale modeling stages: (1) spatial–temporal decomposition by TVGMD; (2) feature selection via functional connectivity; (3) mode forecasting using ML techniques; and (4) forecast ensemble through simple addition. The novelty of this work stems from the application of TVGMD, which treats air quality time series as time-varying graphs and concurrently extracts constituent oscillatory modes along with their connectivity structures (called graph mode) in a fully data-driven manner. This will greatly help to capture spatial–temporal dependencies of the target station's air quality at each timescale. We conduct extensive experimental evaluations on an hourly PM2.5 concentration dataset containing records of 36 air quality monitoring stations in Beijing, China, and empirical results show that the developed methodology outperforms other comparable methods. Overall, this article offers a new perspective and an effective framework for air quality prediction as well as multi-scale spatial–temporal analysis.
AB - Reliable air quality forecasting is a challenging task, as we need to consider both complex temporal and spatial patterns. Following the philosophy of the well-established “decomposition and ensemble” modeling framework, this paper proposes a novel multi-scale spatial–temporal ensemble learning methodology based on time-varying graph mode decomposition (TVGMD) and machine learning (ML) techniques to jointly analyze the temporal dynamics and spatial interactions of multi-station air quality at multiple timescales, aiming to achieve more accurate air quality predictions. The suggested methodology comprises four multi-scale modeling stages: (1) spatial–temporal decomposition by TVGMD; (2) feature selection via functional connectivity; (3) mode forecasting using ML techniques; and (4) forecast ensemble through simple addition. The novelty of this work stems from the application of TVGMD, which treats air quality time series as time-varying graphs and concurrently extracts constituent oscillatory modes along with their connectivity structures (called graph mode) in a fully data-driven manner. This will greatly help to capture spatial–temporal dependencies of the target station's air quality at each timescale. We conduct extensive experimental evaluations on an hourly PM2.5 concentration dataset containing records of 36 air quality monitoring stations in Beijing, China, and empirical results show that the developed methodology outperforms other comparable methods. Overall, this article offers a new perspective and an effective framework for air quality prediction as well as multi-scale spatial–temporal analysis.
KW - Air quality
KW - Graph signal processing
KW - Multi-scale modeling
KW - Spatial-temporal analysis
KW - Time series forecasting
UR - https://www.scopus.com/pages/publications/105007970631
U2 - 10.1016/j.eswa.2025.128538
DO - 10.1016/j.eswa.2025.128538
M3 - Journal article
AN - SCOPUS:105007970631
SN - 0957-4174
VL - 291
JO - Expert Systems with Applications
JF - Expert Systems with Applications
M1 - 128538
ER -