A Comparative Review of Attribute Selection Techniques for PM2.5 Prediction Using Machine Learning Models

Abstract

Accurate prediction of fine particulate matter (PM2.5) is vital for understanding and mitigating air pollution impacts on public health. With the rise of machine learning (ML) in environmental forecasting, selecting the most influential features remains a critical preprocessing step. This review paper evaluates the effectiveness of various attribute selection techniques applied to PM2.5 prediction, including filter, wrapper, and embedded methods. We compare the results from Random Forest, LASSO regression, Recursive Feature Elimination (RFE), and correlation analysis. Our comparative analysis reveals that Random Forest consistently highlights meteorological variables such as temperature and wind speed as top contributors, whereas LASSO reduces model complexity by focusing on core pollutants. The paper provides insights for researchers aiming to develop robust and computationally efficient models for real-time PM2.5 forecasting.