Application of the Lasso regularisation technique in mitigating overfitting in air quality prediction models

Area of Science:

• Environmental Science and Engineering
• Data Science and Machine Learning
• Atmospheric Chemistry and Physics

Background:

• Air pollution poses significant global challenges to public health and ecological sustainability.
• Machine learning (ML) models are increasingly used for air quality forecasting but suffer from overfitting.
• Overfitting reduces the efficacy and generalizability of ML models, necessitating regularization techniques.

Purpose of the Study:

• To enhance the precision of air quality forecasting models by applying the Least Absolute Shrinkage and Selection Operator (Lasso) regularization technique.
• To mitigate overfitting and improve the generalizability of models predicting concentrations of PM2.5, PM10, CO, NO2, SO2, and O3.
• To evaluate the effectiveness of Lasso in feature selection and model performance improvement using Tehran's air quality data.

Main Methods:

• Utilized a comprehensive dataset from 16 sensors in Tehran, Iran, spanning from 2013 to 2023.
• Applied the Lasso regularization technique to machine learning models for air pollutant concentration forecasting.
• Evaluated model performance using R-squared (R2), Mean Absolute Error (MAE), Mean Square Error (MSE), Root Mean Square Error (RMSE), and Normalized Mean Square Error (NMSE) indices.

Main Results:

• Lasso regularization significantly enhanced model reliability by reducing overfitting and identifying key predictive features for particulate matter (PM2.5: R2=0.80, PM10: R2=0.75).
• Forecasting performance for gaseous pollutants (CO, NO2, SO2, O3) remained unsatisfactory (R2 ranging from 0.35 to 0.55), attributed to their high dynamism and complex chemical interactions.
• The strong performance for PM prediction was likely due to minimal missing data, whereas gaseous pollutant models faced challenges inherent to their nature and the chosen model architecture.

Conclusions:

• Lasso regularization is a highly effective technique for mitigating overfitting and selecting important features in air quality forecasting models.
• The study highlights the challenges in accurately forecasting gaseous pollutants compared to particulate matter due to inherent atmospheric complexities.
• The successful application of Lasso suggests its strong potential for broader adoption in developing robust and reliable air quality prediction systems.