Missing Value Imputation of Time-Series Air-Quality Data via Deep Neural Networks

Tae-Sung Kim, Jin-Hee Kim, Wonho Yang, Hunjoo Lee, Jaegul ChooInternational Journal of Environmental Research and Public Health 2021 · DOI

TL;DR

This paper introduces a novel deep learning model for imputing missing values in air quality time-series data. By decomposing the time series into bias, slope, seasonality, and residual components, the model achieves high accuracy while offering enhanced interpretability, a significant improvement over existing black-box deep learning methods.

What Problem Does This Paper Address?

Accurate analysis of air quality data is crucial for understanding and mitigating air pollution. However, real-world air quality monitoring systems often suffer from missing data due to sensor malfunctions or human error. This is particularly challenging for spatio-temporal data collected from multiple locations simultaneously. While deep learning models have shown promise in imputing these missing values, their lack of interpretability makes it difficult to understand why certain values are imputed, hindering trust and further analysis. This paper aims to bridge this gap by developing an interpretable deep learning imputation model that performs comparably to or better than existing methods.

Methods

The proposed model is inspired by the N-BEATS architecture, known for its interpretability in time-series forecasting. It tackles missing value imputation by sequentially decomposing the input time-series into four distinct components:

  • Bias Block: Estimates the average value of the non-missing data.
  • Slope Block: Captures the linear trend of the time series.
  • Seasonality Block: Models the periodic patterns within the data.
  • Residual Block: Accounts for the remaining unpredictable variations.

The model learns to predict these components, and their summation reconstructs the original time series. For imputation, the model predicts these components for the missing time steps. The key novelty lies in explicitly defining the mathematical forms for slope (linear) and seasonality (Fourier series), allowing for a more interpretable breakdown of the imputed values. The model is trained to minimize the mean absolute error between the reconstructed time series and the ground truth.

Datasets

The study utilizes two publicly available air quality datasets:

  • Guro-gu air quality dataset: Collected from 24 locations in Seoul, South Korea, with minute-level data from January 2020 to July 2021.
  • Dangjin-si air quality dataset: Collected from 42 locations in Dangjin-si, South Korea, with minute-level data from May 2020 to July 2021.

For experiments, PM2.5 and PM10 values were used. To evaluate imputation performance, 20% of the test data was artificially made missing, in addition to naturally occurring missing values.

Key Insights and Ideas

  • Interpretability is Achievable in Deep Learning Imputation: The paper demonstrates that deep learning models for imputation don't have to be black boxes. By decomposing the time series into meaningful components, we can understand how the model arrives at its imputed values.
  • Component-wise Imputation is Effective: Separating bias, slope, and seasonality allows the model to learn and represent different aspects of the time series more effectively, leading to better imputation accuracy.
  • Spatio-temporal Information is Leveraged: The model implicitly uses information from multiple locations to improve imputation, acknowledging the interconnectedness of air quality across different areas.
  • N-BEATS Architecture is Adaptable: The success of N-BEATS in forecasting is shown to be transferable to the imputation task, highlighting the flexibility of its interpretable design.

Key Results

The proposed model consistently outperformed baseline methods, including Mean substitution, Spatial Average (SA), and Multivariate Imputation by Chained Equations (MICE), on both datasets.

  • Quantitative Performance:
    • On the Guro-gu dataset, the proposed model achieved MAE of 1.170 and sMAPE of 7.155 for PM2.5, and MAE of 2.738 and sMAPE of 9.385 for PM10. These figures were significantly lower than MICE (MAE: 4.825/9.291, sMAPE: 28.865/31.408) and other baselines.
  • Similar improvements were observed on the Dangjin-si dataset, with the proposed model achieving MAE of 1.149/4.664 and sMAPE of 9.710/13.702 for PM2.5/PM10, respectively.
  • Qualitative Results: Visualizations showed that the proposed model's imputed time series closely matched the original "label" (ground truth) time series, outperforming baselines that struggled to capture the nuances of the data.

Limitations and Future Work

  • Seasonality Modeling: The paper acknowledges that the current seasonality block, using a finite discrete Fourier series, struggles with seasonalities that have periods significantly different from the pre-defined ones. This is particularly problematic for long-period seasonalities (e.g., yearly) which are computationally challenging to handle.
  • Future Directions: The authors plan to explore more appropriate seasonality functions for air quality data and investigate methods to better handle long-period seasonalities. They also emphasize that the interpretability of their model allows for easier identification of such limitations.

Summary updated

Abstract

To prevent severe air pollution, it is important to analyze time-series air quality data, but this is often challenging as the time-series data is usually partially missing, especially when it is collected from multiple locations simultaneously. To solve this problem, various deep-learning-based missing value imputation models have been proposed. However, often they are barely interpretable, which makes it difficult to analyze the imputed data. Thus, we propose a novel deep learning-based imputation model that achieves high interpretability as well as shows great performance in missing value imputation for spatio-temporal data. We verify the effectiveness of our method through quantitative and qualitative results on a publicly available air-quality dataset.
BibTeX
@article{kim2021missing,
  author = {Tae-Sung Kim and Jin-Hee Kim and Wonho Yang and Hunjoo Lee and Jaegul Choo},
  title = {{Missing Value Imputation of Time-Series Air-Quality Data via Deep Neural Networks}},
  journal = {International Journal of Environmental Research and Public Health},
  year = {2021},
  doi = {10.3390/ijerph182212213},
}
Figure