Abstract
Many sensors have been deployed in the physical world, generating massive geo-tagged time series data. In reality, readings of sensors are usually lost at various unexpected moments because of sensor or communication errors. Those missing readings do not only affect real-time monitoring but also compromise the performance of further data analysis. In this paper, we propose a spatio-temporal multi-view-based learning (ST-MVL) method to collectively fill missing readings in a collection of geosensory time series data, considering 1) the temporal correlation between readings at different timestamps in the same series and 2) the spatial correlation between different time series. Our method combines empirical statistic models, consisting of Inverse Distance Weighting and Simple Exponential Smoothing, with data-driven algorithms, comprised of User-based and Item-based Collaborative Filtering. The former models handle general missing cases based on empirical assumptions derived from history data over a long period, standing for two global views from spatial and temporal perspectives respectively. The latter algorithms deal with special cases where empirical assumptions may not hold, based on recent contexts of data, denoting two local views from spatial and temporal perspectives respectively. The predictions of the four views are aggregated to a final value in a multi-view learning algorithm. We evaluate our method based on Beijing air quality and meteorological data, finding advantages to our model compared with ten baseline approaches.
Introduction
In the era of "smart cities," sensor networks are critical for monitoring environmental factors like air quality and weather. However, these networks are inherently fragile, frequently suffering from data gaps due to power outages, communication failures, or hardware malfunctions. Filling these gaps is challenging due to "block missing data"—where sensors are offline for extended periods or entire regions lose connectivity—and non-linear correlations, where physical proximity does not always guarantee similar readings.
As shown in Figure 1, missing readings can occur across both space and time. Traditional interpolation or matrix factorization methods often struggle when entire rows or columns of data are absent, making it difficult to maintain real-time monitoring accuracy.
Methodology: The ST-MVL Framework
To address these challenges, we propose ST-MVL (Spatio-Temporal Multi-View Learning). The framework integrates four distinct perspectives to capture the complex dynamics of geo-sensory data.
The architecture, illustrated in Figure 3, consists of four views:
- Global Spatial View (IDW): Uses Inverse Distance Weighting, assuming that closer sensors generally provide more relevant information over the long term.
- Global Temporal View (SES): Employs Simple Exponential Smoothing based on long-term historical trends of a single sensor.
- Local Spatial View (UCF): Adapts User-based Collaborative Filtering to identify "neighbors" with similar recent behavior, regardless of physical distance.
- Local Temporal View (ICF): Adapts Item-based Collaborative Filtering to capture instantaneous trends within a recent time window.
These views are combined using a multi-view learning algorithm that learns optimal weights for each prediction by minimizing the error against ground truth data, allowing the system to dynamically prioritize the most trustworthy view for any given missing value.
Experimental Evaluation
The ST-MVL model was evaluated using two real-world datasets from Beijing: air quality data (36 stations) and meteorological data (16 sensors) recorded hourly from May 2014 to April 2015.
The results demonstrate that ST-MVL significantly outperforms traditional baselines such as Kriging, SARIMA, and Non-negative Matrix Factorization (NMF). For PM2.5 data, ST-MVL achieved a Mean Relative Error (MRE) of 17.4%, compared to 23.4% for Kriging. Notably, the model excels in handling sudden changes and block missing scenarios where temporal or spatial continuity is broken.
Conclusion and Future Work
ST-MVL provides a robust solution for data imputation in sensor networks by balancing global empirical assumptions with local data-driven insights. Future research will explore non-linear aggregation methods, such as deep learning, and the integration of external features like traffic patterns and points of interest to further refine local context modeling.
Reference: Yi, X., Zheng, Y., Zhang, J., & Li, T. (2016). ST-MVL: Filling Missing Values in Geo-Sensory Time Series Data. Proceedings of the Twenty-Fifth International Joint Conference on Artificial Intelligence (IJCAI-16).
Summary updated